TobyOnFitnessTech
←All Episodes
Claude Opus 5.5 Lands at $4 In, $20 Out, Meets GPT-6 Sol and Astra — Episode 116 cover art
Episode 116·September 24, 2026·33:40

Claude Opus 5.5 Lands at $4 In, $20 Out, Meets GPT-6 Sol and Astra

Claude Opus 5.5 lands at $4 input and $20 output with Fable-class results, facing GPT-6 Sol and Astra in this episode's head-to-head. Show notes: https://tobyonfitnesstech.com/podcasts/episode-116/

🎧 Listen to Episode

AgentStack Daily EP116 — Opus 5.5 ships at $4/$20, Grok 4.7 lands, Nemotron tracks eight speakers

Title: Claude Opus 5.5 Lands at $4 In, $20 Out, Meets GPT-6 Sol and Astra

Tagline: Claude Opus 5.5 ships at $4 input and $20 output with what Anthropic calls Fable-class results, landing directly into a head-to-head with OpenAI's GPT-6 Sol and Astra. Grok 4.7 also arrives as a larger model at the same $2 and $6 price point. Hermes Agent v2026.9.21 headlines this week's harness readout. NVIDIA's Nemotron 3 diarization tracks eight overlapping speakers live, and Nokia open-sources AnyJev, a no-training layer that turns open LLMs into reliable decision-makers. OpenAI extends Daybreak cyber access to Ukraine and releases MentalHealthBench for mental-health conversations. The research digest covers GameHorizon's 5,000-hour AAA agent test and a new benchmark for agent taste on long tasks.

Feed description: Claude Opus 5.5 lands at $4 input and $20 output with Fable-class results, facing GPT-6 Sol and Astra in this episode's head-to-head. Grok 4.7 ships at $2 and $6 as a larger model. Hermes Agent v2026.9.21 leads the harness readout. NVIDIA's Nemotron 3 diarization handles eight overlapping speakers live, and Nokia open-sources AnyJev for reliable open-LLM decision-making. OpenAI extends Daybreak cyber access to Ukraine, releases MentalHealthBench, and the research digest covers GameHorizon's 5,000-hour AAA agent benchmark plus a new study on agent taste across long tasks.

Story Slate

  1. Agent Stack Release Readout: Hermes Agent v2026.9.21 Hermes Agent shipped v0.21.4, tagged v2026.9.21, on September 21, 2026. It is a patch release that rolls roughly 1,800 merged pull requests and 5,071 non-merge commits across 5,169 changed files into a stable cut for Docker, Hermes Cloud, and hosted deployments. Curated release notes for the window are deferred to v0.22.0. Notable additions in the window include --format stream-json for the CLI, an mcp.discovery_concurrency cap, skills.auto_load pinning skills into every new session, and a host-wide gateway singleton lock so Desktop attaches to a running backend instead of spawning a second one. Technical depth angle: The notable mechanism is the host-wide gateway singleton: one backend process owns connector operations and serves Desktop, TUI, and CLI through a rendezvous record, eliminating duplicate processes. Stream-json output and the MCP discovery concurrency cap give operators structured logs and a tunable upper bound on parallel MCP connections during cold start. Actionability angle: What this means is that self-hosters running Docker images or Hermes Cloud get a stable, tagged build to pin against instead of tracking main. The skills auto-load behavior and stream-json CLI output change how new sessions start and how downstream tooling can ingest runs. Watch v0.22.0 for the deferred curated notes covering this whole window. The implications land across Antigravity, Codex, Claude Code, and Hermes stacks alike. Listener hook: If you pin Hermes Agent by Docker tag or live in Hermes Cloud, this is the build you're moving to today.

  2. Claude Opus 5.5 delivers Fable-class results at $4 in, $20 out Anthropic released Claude Opus 5.5 on September 22, positioning it between flagship and mid-tier. The model performs about as well as Claude Fable 5.1 on most benchmarks while costing 40% less to run at default settings than Opus 5. Input pricing dropped 20% to $4 per million tokens, output to $20, and cache reads are 60% cheaper. Output generates 30% faster than Opus 5. The thinking model can no longer be disabled, and a Fast mode delivers 2.5x speed at double the price. It runs on Claude's platform, AWS, Google Cloud, Azure, and GitHub Copilot. Technical depth angle: Opus 5.5 uses adaptive thinking with five effort levels—low, medium, high, xhigh, and max—where higher effort produces more reasoning steps before output. Benchmarks at max effort show meaningful gains: Terminal-Bench 4.0 reaches 66.4% versus Opus 5's 52.3%, AutomationBench hits 40.0% versus 26.9%, and OSWorld partial scores 81.8% versus 74.0%. The thinking model is now mandatory; turning it off is no longer an option. Actionability angle: This means you can run Fable-class code migrations and audits at Opus pricing—Anthropic cites a 680,000-line code migration finishing in under a day and a 200,000-line audit completing in under three hours versus over twenty hours on Opus 5. Stripe had 40 stacked pull requests pass CI, Deloitte caught 72% of bugs at low effort versus 56% for Opus 5, and Box cut token usage by a third. If speed matters more than cost, Fast mode runs 2.5x faster at $8 input and $40 output per million tokens. Listener hook: If you've been paying Opus 5 prices, Opus 5.5 delivers Fable-class performance for 40% less—and Anthropic has published the full benchmark table so you can verify the claims yourself.

  3. Head-to-head: GPT-6 Sol and GPT-6 Astra versus Anthropic's new flagship Anthropic shipped Claude Opus 5.5 on September 22, 2026, and OpenAI followed roughly 90 minutes later with GPT-6 Sol and Luna. The pricing picture is now clear: Opus 5.5 lists at $4 per million input tokens and $20 per million output, versus GPT-6 Astra's $10 and $50. On benchmarks Anthropic published alongside the launch, Opus 5.5 at its default effort level ties Astra on Terminal-Bench, beats Astra's best on FrontierCode at about one-fifth the cost, and leads on Humanity's Last Exam. Astra retains edges on science and automation tasks. All three models share 128,000-token maximum output and support text, image, and file input. OpenAI positioned Sol as a coding-focused alternative to Astra at $2 and $10, with Luna targeting high-volume work at a fraction of a cent per token. Technical depth angle: Opus 5.5 benchmarks at medium effort tie or beat Astra on several tests while costing 40–80% less per task. Astra leads on AutomationBench and Terminal-Bench-Science. Sol and Luna sit two and four price tiers below Astra, with Luna scoring comparable to Opus 5 at medium effort at 93% less per task. Actionability angle: For tasks in coding, general reasoning, and complex professional work, Opus 5.5 now offers a cost-per-quality point that challenges Astra's flagship positioning. The competitive pricing pressure from both Anthropic and OpenAI means teams running high inference volumes should re-evaluate whether the premium tier is justified for their workload. Sol targets the same coding use case at even lower cost, creating a new evaluation choice between Anthropic's model family and OpenAI's segmented offerings. Listener hook: Anthropic's new flagship undercuts OpenAI's top model by 60% while beating it on most of the same benchmarks.

  4. Grok 4.7 Lands: Bigger Model, Same $2/$6 Price SpaceXAI shipped Grok 4.7 on September 21, 2026, billing it as the company's most capable model for coding and knowledge work. The new release sits on a larger base model than Grok 4.6 and was trained with a longer reinforcement learning run weighted toward multi-hour problems. Pricing holds at $2 per million input tokens and $6 per million output tokens — unchanged from Grok 4.6 — with an optional fast variant that doubles output speed for twice the price. Grok 4.7 sits at the frontier of CursorBench 4.0 price-performance, tops LatchBio's biosafety benchmark at 62.4%, and is live now in Cursor, the Grok API, and Grok Build. Technical depth angle: The upgrade is a new, larger base model trained with a longer RL run on a harder multi-hour task mix, plus native understanding of the Grok Bot harness for conversational work. The serving stack — same $2/$6 pricing, same speed, optional 2x-speed fast variant — stayed constant, so the wins come from training rather than new infrastructure. Actionability angle: For builders, Grok 4.7 is a drop-in upgrade at the same price, so existing Grok API code keeps working while long-context coding and document generation both improve. It's worth pitting against your current coding harness on multi-hour refactors and professional workflows like legal and financial analysis. Listener hook: SpaceXAI's Grok 4.7 ships a bigger brain at the same $2/$6 price as its predecessor — and tops the biosafety leaderboard while doing it.

  5. Nokia Open-Sources AnyJev: No-Training Layer Makes Open LLMs Reliable Decision-Makers Nokia's applied research team has open-sourced AnyJev, a Python library that transforms open-source language models into calibrated decision systems without requiring additional training. The library installs via PyPI under Apache-2.0 licensing and supports multiple backend frameworks including transformers and vLLM. AnyJev operates by rotating option presentations and averaging token probabilities across multiple passes to eliminate positional and label bias, then applies temperature scaling for additional calibration. In benchmarks on Qwen3-8B, the system reduced decision flip rates by 68% when options were reordered, boosted accuracy by 8 percentage points, and increased the proportion of automatically resolvable queries from under 8% to over 52%. Technical depth angle: AnyJev operates in two calibration layers. L0 uses cyclic shifts and batch prior correction to remove position and label bias without any labeled data. L1 adds temperature scaling using 100-500 labeled examples to reshape confidence estimates. The core mechanism is presenting each decision in K rotations and averaging the log probabilities—geometric mean in log space—then dividing out the observed prior distribution at strength 0.75. This removes additive position bias exactly and produces probability scores suitable for thresholding. Actionability angle: For builders deploying open models in production, AnyJev provides a straightforward path to turn any transformer or vLLM-served model into a decision system for routing, classification, or structured selection tasks. The library is immediately usable via PyPI with vLLM prefix caching for efficient serving, requiring no fine-tuning or data collection pipeline. Listener hook: Nokia is open-sourcing a technique that eliminates the two most common pitfalls when reading probabilities directly from language models, making open-source decisions reliable enough for production traffic.

  6. NVIDIA's Nemotron 3 Diarization Tracks Eight Overlapping Speakers Live NVIDIA has released Nemotron 3 Diarization, a 100-million-parameter open-weight model that answers one question about any conversation: who spoke when. It tracks up to eight speakers, including when voices overlap, in a single checkpoint that covers offline transcripts down to 320-millisecond streaming. On Voice Arena's initial Diarization-Bench, the model hit 14.72% error rate versus 19.3% for the next-ranked system, a roughly 24% relative reduction. Against NVIDIA's prior four-speaker baseline at 1.04 s latency, average error dropped 41% across eight evaluation conditions. Weights ship under the OpenMDW 1.1 license for commercial use and run through NeMo on Ampere, Ada Lovelace, Hopper, or Blackwell GPUs. Pair it with an ASR model like Parakeet TDT 0.6B and you have a speaker-attributed transcript pipeline. Technical depth angle: The model converts 16 kHz audio into mel-spectrogram frames with a 10 ms step, then stacks every eight frames into 80 ms encoder tokens. A 31-layer Transformer with rotary positional embeddings processes those tokens, and a one-dimensional convolution up-samples predictions back to 10 ms resolution, so overlapping voices register as multiple active channels in an eight-channel tensor. Speakers are ordered by arrival time, with streaming memory held by an Arrival-Order Speaker Cache and a first-in-first-out queue to keep labels stable across chunks. Actionability angle: What this means: speaker-attributed transcripts are now within reach of any team willing to run a 100M-parameter checkpoint on NVIDIA hardware. Production deployments can use Baseten or DigitalOcean, and on-device mobile is available through Argmax Pro SDK 3. Why this matters: meeting tools, call analytics, podcast pipelines, and voice-agent memory can stop relying on guesswork for who said what. Listener hook: NVIDIA just dropped a 100-million-parameter open model that finally tells you who was talking in any recording.

  7. OpenAI extends Daybreak cyber access to Ukraine OpenAI announced on September 23, 2026 that it is extending access to its Daybreak program to the Government of Ukraine to support the cyber defense of civilian infrastructure. The announcement, published via OpenAI News, frames the extension as part of the company's broader work helping protect civilian-facing systems from cyber threats. Daybreak is positioned as OpenAI's program for cyber defense support, with Ukraine becoming a direct government recipient of extended access. Technical depth angle: Daybreak is OpenAI's program supporting cyber defense for civilian infrastructure. Extending access to a national government means OpenAI is shipping direct vendor-to-government access for civilian-defense use, treating civilian infrastructure as a distinct category that warrants dedicated tooling and program structure rather than ad-hoc support. Actionability angle: What this means is that direct vendor-to-government access programs for civilian cyber defense are an active model rather than one-off partnerships. For builders working on threat-intel pipelines, anomaly detection, or automated triage for critical-infrastructure buyers, the takeaway is that the civilian-defense buyer is real and active. The next thing worth watching is how Daybreak tooling gets used in practice inside Ukraine's defensive stack. Listener hook: Ukraine's government is getting extended access to OpenAI's Daybreak cyber program — here's what that actually changes for civilian defense.

  8. Research digest: A benchmark for agent 'taste' on long tasks A research team has released TasteBench, a benchmark aimed at measuring the in-the-middle decisions an AI agent makes while working through a long task — not just whether the run ends in success. The authors frame the quality of those mid-flight choices — which hypothesis to test, which implementation path to follow — as the agent's "taste." End-to-end benchmarks already exist, but this one focuses on the branching decision points themselves. It could matter because real software and research work is full of these choices, and right now the field has no clean way to compare how well different agents make them. Technical depth angle: The paper proposes measuring an agent's decision quality at choice points during long-running work — what it picks to test, pursue, or abandon — rather than only scoring final outputs. Actionability angle: What this means for teams building long-running agents: there is now a separate yardstick for decision quality inside a run, not just whether the task finished cleanly. Why this matters: it surfaces failure modes that show up only when an agent picks the wrong starting branch and then runs with it for hours. Listener hook: Most agent tests count a win or a loss — this one tries to grade the judgment calls along the way.

  9. Research digest: GameHorizon Puts AI Agents Through 5,000 Hours of AAA Gameplay A new benchmark suite tests AI models on 5,000 hours of AAA video game play, covering 21 titles from Valorant to Elden Ring. The goal: measure how well models can plan across different time horizons — from a single keystroke to a multi-session strategy — using language-grounded instructions. The team tested 47 different models and built offline multiple-choice questions plus stepwise online tasks that pinpoint exactly where a model fails. Technical depth angle: The suite gives AI agents a pyramid of instructions — primitive actions, short-horizon operations, medium-horizon goals, and long-horizon strategies — tied to timestamped human gameplay, so evaluators can test planning at every scale against reproducible offline questions and stepwise online tasks. Actionability angle: For anyone building game agents or evaluating multimodal models, this offers a standardized yardstick that isolates where breakdowns happen — perception, planning, or execution. Researchers can compare model families on the same AAA content offline rather than relying on flaky live rollouts. The 62 online subtasks make it possible to localize exactly which step a model gets wrong. Listener hook: If you've ever wondered whether an AI can actually play a real game — not a simplified grid world — here's the first benchmark built to find out.

  10. NVIDIA Frames AI Agent Security as an Engineering Discipline On September 21, NVIDIA published a blog arguing that AI agent security is fundamentally an engineering problem requiring defined requirements, enforceable controls, named owners, and evidence that protections actually work. The post breaks the agent stack into models, harnesses, and runtimes, each carrying distinct security responsibilities, and points to concrete tools — NVIDIA's open-source OpenShell sandbox, Cisco's DefenseClaw governance layer, JFrog's skill scanning, Palo Alto Networks' Prisma AIRS for red teaming, and Capital One's VulnHunter for code security — that defenders can adopt today. Technical depth angle: The piece reframes agent security around three stack layers: models provide capabilities, harnesses organize context and tools, and runtime environments host execution. Crucially, controls must live outside the agent's reach so boundaries hold even when the model's reasoning goes wrong. Actionability angle: For builders shipping agents, this means treating the runtime sandbox — not the prompt — as the actual security perimeter. Wire scoped identities, network policies, and protected audit logs into the execution environment first, and assign a named owner who signs off on reproducible test evidence before each deployment. The framework matters because prompt-level guardrails alone cannot stop an agent with file and network access from doing real damage. Listener hook: If your agent can read a file and call an API, it already has the power to exfiltrate data, and the prompt is not where you stop it.

  11. Thinking Machines Lab moves open-model inference to Crusoe Cloud under $65M deal Thinking Machines Lab has signed a $65 million annual deal to run production inference for its open models on Crusoe Cloud, the companies announced on September 23. The workload runs through Crusoe Managed Inference, a managed serving layer the company tunes around throughput, price-performance, and reliability — and the deal positions Crusoe as a serious neocloud option for open-model builders. Technical depth angle: Crusoe Managed Inference is a managed serving layer that lets a lab point open-model traffic at Crusoe's stack without owning the underlying hardware. The partnership is framed around three operational targets: throughput, price-performance, and reliability. Actionability angle: Builders using Thinking Machines Lab's open models should expect the same model behavior but a different provider behind the endpoints, which is worth re-benchmarking once traffic shifts. If you're running cost-sensitive batch jobs, a managed SLA and explicit price-performance framing may matter more than peak tokens-per-second claims. Why this matters: serving-stack choices now show up directly in your latency, uptime, and per-token bill. Listener hook: A serious open-model lab just bet $65 million a year that Crusoe can serve its traffic more reliably than its previous setup.

  12. Jev returns probabilities instead of text, undercutting GPT-5 Nano on price TypeSafe AI has unveiled Jev, a new model category it's calling 'System One' or 'decision' models. Unlike a regular LLM, Jev takes in text and returns floating point numbers — confidence scores, yes/no probabilities, or distributions over a set of choices — rather than freeform text. It's also priced unusually: only input tokens count, output is free, and the input rate is $0.042 per million tokens, undercutting OpenAI's GPT-5 Nano at $0.05. That makes it a different kind of building block — better suited to classification, scoring, and routing decisions than to writing prose. Technical depth angle: Jev accepts a 'state' object — a string, list of strings, or set of name-value pairs — plus one or more typed questions, and returns a structured probability answer for each. Three question shapes: Noul (yes/no, returning a 0–1 confidence; the name comes from Bernoulli), Choice (a probability distribution over labeled options), and Score (a numeric rating along a sequence of levels). Actionability angle: For builders, this is a fit for routing, triage, and classification layers where a confidence number matters more than a sentence. It pairs naturally with a regular LLM — Jev makes the decision, a text model drafts the reply. Listener hook: Sometimes you don't want an answer, you want a probability — and a new model class just made that cheaper than ever.

  13. A Self-Hosted Browser That Lets AI Agents Skip Captchas A new release of invisible_playwright_mcp, an open-source Python Model Context Protocol server, lets AI agents drive an undetected stealth Firefox for scraping, browser automation, and computer use without running into captchas. The project shipped v0.70.0 on September 23 and now sits at over 31,600 stars on GitHub, making it one of the most-starred stealth-browser tools available to builders. Technical depth angle: It's an MCP server wrapping Playwright around a patched, anti-detect Firefox build that mimics a normal user session, so the traffic profile avoids the fingerprint and bot-detector checks that typically trip up headless browsers. Actionability angle: Builders whose agents need to log in, scrape, or fill forms can self-host this instead of paying for a captcha-solving service. The setup is open-source and runs inside your own environment, which keeps credentials off third-party endpoints. Worth pinning the version and keeping usage inside site terms and applicable law, since stealth browsers live in an active arms race. Listener hook: If your agent keeps getting stopped by captchas, there's a self-hosted stealth Firefox that wants to fix it.

  14. OpenAI releases MentalHealthBench to test AI responses in mental-health conversations On September 23, OpenAI published MentalHealthBench, a new evaluation set designed to measure how helpful and safe AI systems are when responding to realistic mental-health conversations. The benchmark is described as expert-informed, meaning mental-health specialists helped shape the scenarios and the criteria used to score AI replies. OpenAI framed it as a resource for developers and researchers building conversational AI in sensitive domains. Technical depth angle: MentalHealthBench is built from realistic mental-health conversation scenarios shaped with input from mental-health specialists, and it grades AI replies on both helpfulness and safety at the same time so neither dimension can be optimized at the expense of the other. Actionability angle: If you build a chatbot, assistant, or model that touches emotional support, counseling, or any crisis-adjacent use case, this gives you a public yardstick to track regressions and compare options on the same scenarios rather than cherry-picked demos. The dual helpful-and-safe framing also signals what OpenAI considers table stakes for mental-health-facing products, which is useful context for anyone shipping in that space. Listener hook: OpenAI just put a public scorecard on the table for AI in therapy-style conversations, and the rubric came from the people who do this work for real.

Editorial Mix Check

  • flagship_products: 7
  • builder_projects: 6
  • local_ai: 2
  • hardware_compute: 3
  • policy_regulation: 5
  • research: 2

Model Discovery Check

  • Z.ai: GLM 5.3 Prime (z-ai) — Newly listed this cycle (verified September 23, 2026). Primary source: https://openrouter.ai/models/z-ai/glm-5.3-prime. Availability: API via OpenRouter. Capabilities: context length 1000000; GLM-5.3-Prime is the high-speed variant of Z.ai's GLM-5.3, inheriting its full capabilities while delivering 1.5–2× the output throughput through inference acce. Try now / integration angle: available for evaluation via the model page above. Decision: Not Selected — variant/duplicate of a model featured on a recent broadcast, or not a major standalone drop.

  • Qwen: Qwen3.8 Max Prime (qwen) — Newly listed this cycle (verified September 23, 2026). Primary source: https://openrouter.ai/models/qwen/qwen3.8-max-prime. Availability: API via OpenRouter. Capabilities: context length 1000000; Qwen3.8 Max Prime is a higher-throughput variant of Qwen3.8 Max from Alibaba's Qwen team, served as a separate SKU at a higher price point. It accepts text, ima. Try now / integration angle: available for evaluation via the model page above. Decision: Not Selected — variant/duplicate of a model featured on a recent broadcast, or not a major standalone drop.

Local LLM Spotlight

  • prism-ml/Ternary-Bonsai-2-27B-gguf — https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf — Trending open model on Hugging Face; task text-generation; 1950 likes and 2815979 downloads. Tags: llama.cpp, gguf, ternary, 2-bit, llama-cpp, cuda, metal, on-device, hybrid-attention, prismml. Try now: Read the linked model card before downloading, and choose a runtime or device only after it confirms the license, weight format, context window, benchmarks, and hardware requirements.

GitHub Project Radar

  • HKUDS/nanobot — https://github.com/HKUDS/nanobot — nanobot is an ultra-lightweight, open-source, self-hosted Python framework for personal AI agents, shipping a WebUI, tool registry, memory, MCP integration, multi-agent workflows, and automation hooks. stars: 48,518; stars_delta_30d: +1,267 (+2.7%) since 2026-08-21; latest_release: v0.3.5 (2026-09-15). Why this is on the radar now: v0.3.5 shipped on 2026-09-15 and the repository was updated on 2026-09-23. Stack improvement angle: It can sit beside OpenClaw or Hermes as a self-hosted orchestration shell, offloading WebUI, memory persistence, and MCP tool wiring from the harness itself. Try now: Clone the repo and run its WebUI locally to compare its MCP tool surface against what your current harness exposes.

  • DeusData/codebase-memory-mcp — https://github.com/DeusData/codebase-memory-mcp — codebase-memory-mcp is a high-performance MCP server that builds a persistent knowledge graph of any codebase, claiming sub-millisecond queries across 158 languages and a 99% token reduction, shipped as a single static binary. stars: 44,542; stars_delta_30d: +4,787 (+12.0%) since 2026-08-21; latest_release: v0.11.0 (2026-09-15). Why this is on the radar now: v0.11.0 shipped on 2026-09-15 and the repository was updated on 2026-09-23. Stack improvement angle: Drop it into a Codex or Claude Code stack as a dedicated code-intelligence MCP so the agent can query a pre-built graph instead of re-reading files every turn. Try now: Point the static binary at one of your own repos and time a query against the harness's default file-reading path.

  • feder-cr/invisible_playwright_mcp — https://github.com/feder-cr/invisible_playwright_mcp — invisible_playwright_mcp is a self-hosted Python MCP server that drives an undetected anti-detect Firefox profile, giving agents a browser surface that aims to avoid captchas. stars: 31,644; stars_delta_30d: n/a — first tracked appearance; latest_release: v0.70.0 (2026-09-23). Why this is on the radar now: v0.70.0 shipped on 2026-09-23 and the repository was updated on 2026-09-23. Stack improvement angle: Swap it into an OpenClaw or Hermes browser tool to keep scraping and computer-use flows running without the agent hitting captcha walls mid-task. Try now: Stand it up locally and run a single navigation request through your agent to confirm the MCP handshake before wiring it into production flows.

Extra Research Candidates

  • Sam Altman’s remarks at the United Nations Security Council — https://openai.com/index/sam-altman-un-security-council-remarks — OpenAI CEO Sam Altman discusses AI safety, human control, and international cooperation in remarks to the United Nations Security Council. Technical depth angle: The remarks center on AI safety governance, retaining human control over advanced systems, and international coordination frameworks for frontier model deployment.

  • Introducing MentalHealthBench — https://openai.com/index/introducing-mentalhealthbench — MentalHealthBench is an expert-informed benchmark for evaluating helpful and safe AI responses across realistic mental health conversations. Technical depth angle: MentalHealthBench evaluates helpfulness and safety using expert-informed rubrics applied to realistic mental-health conversation transcripts.

  • TykTechnologies/tyk — Open Source API and AI Gateway supporting REST, GraphQL, TCP, gRPC and MCP (Mode — https://github.com/TykTechnologies/tyk — Open Source API and AI Gateway supporting REST, GraphQL, TCP, gRPC and MCP (Model Context Protocol) GitHub reports 10830 stars. Latest release: v5.14.0 2026-07-07T17:34:16Z. Repository pushed 2026-09-23T15:02:32Z. Technical depth angle: Tyk acts as a unified gateway that proxies REST, GraphQL, TCP, gRPC, and MCP traffic behind a single routing and policy layer.

Show Notes

Episode 116 — September 23, 2026

[00:00] Episode hook

Hermes Agent shipped v0.21.4, tagged v2026.9.21, on September 21, 2026, rolling roughly 1,800 merged pull requests and 5,071 non-merge commits across 5,169 changed files into one patch release. Anthropic released Claude Opus 5.5 on September 22 at $4 per million input tokens and $20 per million output tokens, performing about as well as Claude Fable 5.1 across most benchmarks while running 40% cheaper to operate. OpenAI followed roughly ninety minutes later with GPT-6 Sol and Luna. SpaceXAI also shipped Grok 4.7 on September 21, 2026, built on a larger base model than Grok 4.6 with coding and knowledge-work improvements at the same $2 input and $6 output pricing. Nokia open-sourced AnyJev, a Python library turning open-source LLMs into calibrated decision systems without additional training, while NVIDIA released Nemotron 3 Diarization, a 100-million-parameter open-weight model that tracks up to eight overlapping speakers live.

[02:00] Agent Stack Release Readout: Hermes Agent v2026.9.21

Hermes Agent shipped v0.21.4, tagged v2026.9.21, on September 21, 2026. It is a patch release from Nous Research that rolls up roughly 1,800 merged pull requests since v0.21.3 into a single stable tag for downstream consumers like Docker images, Hermes Cloud, and hosted deployments. Measured at commit 4b8a8134009a8727a289bcabeb0019fedd353128, the window since v0.21.3 contains 5,071 non-merge commits across 5,169 changed files, with 312,961 lines added and 62,855 removed, alongside 1,812 merged PRs and 2,116 closed issues. Curated release notes for everything in this window are deferred to v0.22.0.

Several concrete changes landed in the window. The CLI gains `--format stream-json` for structured JSONL output, which downstream tooling can ingest directly. A new `mcp.discovery_concurrency` setting caps parallel MCP discovery connections, giving operators a tunable knob during cold start. `skills.auto_load` now pins selected skills into every new session prompt, so the set of available skills is consistent across sessions without re-invocation. A host-wide gateway singleton lock with a rendezvous record means the Desktop client now attaches to the running host backend rather than spawning a second one, and one backend-owned connector operation shows up as a setup card on Desktop, TUI, and CLI.

Other additions include a `decline` unauthorized-DM behavior for the gateway, `session_search` after and before bounds with OR-relaxed recall retry, and a `hermes sessions set-journal-mode` command. Desktop picks up a chat and UI font picker, one-click local engine updates, and plugin uninstall from the Plugins hub. The video catalogs add LTX 2.5 and Kling O3, and the website now has a page for every catalog plugin and author with pinned-commit READMEs and added or updated sorting. About a dozen community plugins joined the catalog, including tailscale, ssh, shodan, terminal, rss, resetwatch, done-bell, kiwi, cognee, and octen.

To update, git installs run `hermes update`, and Docker or Hermes Cloud users pull from `nousresearch/hermes-agent:v2026.9.21`.

[03:10] Claude Opus 5.5 delivers Fable-class results at $4 in, $20 out

Anthropic shipped Claude Opus 5.5 on September 22, the first model in the Claude 5.5 family, with Sonnet 5.5 and Haiku 5.5 arriving within weeks. The headline claim: Opus 5.5 performs about as well as Claude Fable 5.1 on most work while running at lower cost. The pricing reflects that shift. Input tokens cost $4 per million versus Opus 5's $5, output tokens cost $20 versus $25, and cached reads are 60% cheaper at $0.20. At default settings, typical workloads cost 40% less overall. Output generates more than 30% faster than Opus 5. The model carries a 1,000,000-token context window and supports up to 128,000 output tokens, handling text, image, and file inputs. It is available on the Claude Platform, AWS, Google Cloud, Microsoft Azure, and GitHub Copilot, with zero data retention as an option.

The thinking model is now mandatory. Users can choose effort levels—low, medium, high, xhigh, or max—but cannot disable thinking entirely. Max effort unlocks adaptive thinking, which produces more internal reasoning steps before final output. Fast mode in Claude Code and the Claude Platform runs up to 2.5 times faster at $8 input and $40 output per million tokens, trading cost for speed.

Anthropic published benchmark results comparing Opus 5.5, Fable 5.1, and Opus 5 at max effort with adaptive thinking. On Terminal-Bench 4.0, Opus 5.5 scored 66.4% versus Opus 5's 52.3%. On AutomationBench, the gap widened to 40.0% versus 26.9%. On OSWorld 2.0 partial, Opus 5.5 reached 81.8% versus Opus 5's 74.0%. GDPval-AA v2.1 Elo placed Opus 5.5 at 1846 versus Opus 5's 1708. Anthropic also shared internal test results from customers. Stripe ran 40 stacked pull requests through the model and all passed CI. Box used a third of the tokens compared to previous runs with 40% less verbose output. Deloitte caught 72% of planted bugs at low effort, compared to 56% for Opus 5. Hebbia achieved 86.6% rubric coverage on a research task versus 60.3% with the prior model.

Safety testing by METR and Frontier Design before release showed Opus 5.5 posted Anthropic's best score yet on an automated behavioral audit of roughly 2,000 scenarios and attempted to cross containment boundaries about 85% less often than Opus 5. Anthropic notes the model often suspects it is being evaluated, which tempers confidence in that figure. Cyber tasks route to Opus 4.8 under safeguards, and Anthropic opened a Cyber Verification Program and Life Sciences Verification Program to vetted organizations. Five-hour usage limits are rising for Pro, Max, Team, and seat-based Enterprise plans. Output carries text watermarking for EU AI Act compliance, and preserved thinking applies to Fable 5.1 and Opus 5.5 on API accounts created on or after August 31, 2026.

[05:58] Head-to-head: GPT-6 Sol and GPT-6 Astra versus Anthropic's new flagship

Anthropic shipped Claude Opus 5.5 on September 22, 2026. OpenAI answered roughly 90 minutes later with GPT-6 Sol and Luna, though Astra remains OpenAI's top-tier flagship. The pricing picture is now concrete. Opus 5.5 lists at $4 per million input tokens and $20 per million output. Astra sits at $10 and $50. Sol comes in at $2 and $10, and Luna at a fraction of a cent. On benchmarks Anthropic published with the launch, Opus 5.5 at its default medium effort ties Astra's Terminal-Bench score, beats Astra's best on FrontierCode at roughly one-fifth the cost, and leads on Humanity's Last Exam with tools. Astra retains edges on science-focused and automation benchmarks. All three models share 128,000-token maximum output, text-and-image-and-file input, and context windows near one million tokens. OpenAI positioned Sol as a coding and professional-work alternative to Astra at significantly lower cost, claiming it matches Claude Fable 5.1 performance on FrontierCode and reaches Astra-level factual accuracy at a fraction of the price. Luna, the fastest and cheapest tier, scores on par with Opus 5 at medium effort at 93% less per task, targeting high-volume automated workflows. Simon Willison's read is that the premium top tier is now a two-horse race between Astra and Claude Fable 5.1, while the real price war is happening below that, with Opus 5.5, Sol, and Grok 4.7 competing for cost-conscious builders.

[07:23] Grok 4.7 Lands: Bigger Model, Same $2/$6 Price

SpaceXAI released Grok 4.7 on September 21, 2026, marketing it as the company's most capable model for coding and knowledge work. The twist is the price tag: it lands at the same $2 per million input tokens and $6 per million output tokens as Grok 4.6, with the same serving speed, and an optional fast variant that doubles output speed for twice the price.

Under the hood it's a larger base model than Grok 4.6, trained on a longer reinforcement learning run weighted toward problems that take many hours to complete. That extra training shows up in better self-verification, longer-context handling, and a new instruction layer that natively understands the Grok Bot harness for conversational and general knowledge tasks. SpaceXAI also rebuilt the safeguard stack, claiming Grok 4.7 is the strongest model it has tested on refusals and jailbreak resistance. On LatchBio's biosafety benchmark it posts 62.4%, and on HackerBench v0.3 it lets through only 3.3% of risky dual-use prompts while rarely blocking legitimate security work.

On benchmarks, Grok 4.7 sits at the frontier of CursorBench 4.0 price-performance, with a high-effort score on DeepSWE v1.1 and comparable marks on GDPval, AA Briefcase v1.1, and EEBench against Fable 5.1 and GPT-6 Astra. Document and presentation creation improved over Grok 4.6, and SpaceXAI says select cybersecurity partners now have invite-only access to its red-team capabilities for defense research.

It's live today in Cursor, Grok Build, the Grok API, third-party coding harnesses, and several model routers and cloud platforms. For anyone already running Grok 4.6 in a production pipeline, the upgrade is effectively free at the API level — flip the model string and you get the bigger brain for the same bill.

[09:09] Nokia Open-Sources AnyJev: No-Training Layer Makes Open LLMs Reliable Decision-Makers

Nokia's applied research team has released AnyJev, a Python library that turns any open language model into a calibrated decision model without training. The tool targets a frequent production need: selecting one answer from a fixed set rather than generating free-form text. It installs from PyPI under Apache-2.0 and has transformers and vLLM backends with shared-prefix scoring for efficient serving.

The problem the team addresses is that reading probabilities directly from language models produces two consistent errors. First, models favor certain labels regardless of input—what the team calls prior bias, like preferring "Yes" over "No." Second, models favor positions in the option list, so shuffling the order changes the answer. Both flaws make naive probability reading unreliable for real systems.

AnyJev applies two fixes. L0 uses cyclic shifts: for a question with K options, it shows the list in K different rotations, so every option appears in every position once. It averages the results in log space as a geometric mean, which removes position bias exactly if that bias is additive in logit space. L0 also keeps a running mean of predicted distributions on real inputs and divides it out at strength 0.75 after observing 8 items, correcting the prior bias. L1 adds temperature scaling on top, requiring 100 to 500 labeled examples per question type. That reshapes confidence scores without changing which answer ranks first.

On Qwen3-8B tested against the BANKING77 benchmark—a 20-way classification task with 300 test items—raw logits flipped answers 23% of the time when options were reordered. AnyJev L0 cut that to 7.3%, and L1 held it at 7.7%. Accuracy improved from 74.7% to 80.7%. More practically, the fraction of queries that could be decided automatically at a 5% error threshold rose from 7.7% to 52%, meaning over half of all decisions could be routed with confidence instead of escalated to a human.

The team tested the approach across Qwen, OLMo, Granite, Phi, and Mistral models, with L0 reducing order flips across all nine model and task combinations. On a typed-decisions dataset, Qwen3-32B with L1 reached a calibration error of 0.036 compared to 0.144 reported for Jev, the commercial decision model it emulates.

For serving, developers point vLLM with prefix caching at a hosted model and configure a VLLMBackend in AnyJev. The team estimates roughly 0.25 seconds per decision at batch 32 on a single H100 with K equals 20. The library is available now on PyPI.

[11:40] NVIDIA's Nemotron 3 Diarization Tracks Eight Overlapping Speakers Live

NVIDIA has released Nemotron 3 Diarization, a 100-million-parameter open-weight model that answers one deceptively simple question about any conversation: who spoke when. Automatic speech recognition gives you the words, but without attribution a summarizer cannot tell who made a commitment or who raised an objection. Diarization fills that gap.

The model tracks up to eight speakers and handles overlapping voices in a single checkpoint. NVIDIA's earlier Streaming Sortformer topped out at four. Inputs are 16 kHz, single-channel audio in wav, flac, opus, or mp3 format. A mel-spectrogram with a 10 ms step is stacked by a factor of eight to produce 80 ms encoder frames, which feed a 31-layer Transformer with rotary positional embeddings. A one-dimensional convolution layer upsamples speaker activity back to 10 ms, producing an eight-channel tensor where two voices at once lights up two channels.

Speakers are ordered by arrival time, so the first new voice takes channel one and stays there across streaming chunks. Two memory mechanisms keep that stable: an Arrival-Order Speaker Cache plus a first-in-first-out queue.

On Voice Arena's initial Diarization-Bench, which measures diarization error rate across 139 English conversations totaling about 22 hours, the model hit 14.72% versus 19.3% for the next-ranked system, roughly a 24% relative reduction. Against the four-speaker baseline at 1.04 s latency, average relative error dropped 41% across eight conditions, with a 65% drop on NOTSOFAR1 MHM. Throughput at 30.4 s hit 15,113x real-time on an NVIDIA RTX PRO 5000.

Weights are released under the OpenMDW 1.1 license for commercial use. Production paths include Baseten and DigitalOcean, with on-device mobile through Argmax Pro SDK 3.

[13:21] OpenAI extends Daybreak cyber access to Ukraine

OpenAI announced on September 23, 2026 that it is extending access to its Daybreak program to the Government of Ukraine. The program supports the cyber defense of civilian infrastructure, according to OpenAI News.

The extension puts OpenAI's tooling inside the defensive stack protecting Ukrainian civilian systems from cyber threats. By opening Daybreak access to a national government, OpenAI is treating civilian infrastructure as a category that warrants direct vendor support, rather than routing that support through intermediaries or third-party partnerships.

Daybreak is positioned as OpenAI's program for cyber defense work on civilian-facing systems, and Ukraine becomes a direct government recipient of extended access under that program.

What this changes in practice: any organization building critical-infrastructure defense tools now has another data point showing that major model vendors are willing to ship direct government access programs for civilian cyber work, not just research partnerships or advisory engagements. For builders working on adjacent tooling — threat-intel pipelines, anomaly detection, automated triage, or incident-response copilots — this confirms that the civilian-defense buyer is real, active, and getting direct vendor relationships.

The primary source is OpenAI's newsroom post dated 2026-09-23, published at 13:00 UTC.

[14:32] Research digest: A benchmark for agent 'taste' on long tasks

A research project called TasteBench is asking a fresh question about AI agents that work on long jobs: how good are the decisions they make along the way, not just whether the run ends in success? The paper, currently trending on HuggingFace's daily feed, borrows the word taste to describe an agent's ability to pick the right next move — which hypothesis to chase, which implementation to build on — inside a task that sprawls across hours of work. Existing benchmarks grade the final answer. This one focuses on the branching choices themselves. The authors argue that for software engineering and research workflows, those mid-run choices are what actually decide the outcome. A benchmark that isolates them could let teams compare agents on a dimension they currently have to guess at. Worth watching: whether other labs adopt it, or build their own version of a decision-quality yardstick for long-running work.

[15:29] Research digest: GameHorizon Puts AI Agents Through 5,000 Hours of AAA Gameplay

A team has released GameHorizon, a benchmark and dataset that puts AI agents through 5,000 hours of recorded play across 21 AAA video games — including Valorant, Minecraft, Grand Theft Auto V, Palworld, and Elden Ring. A hundred experienced human players drove the recordings, and every action is timestamped and labeled with a three-level pyramid of language instructions: primitive actions, short-horizon operations, and longer goals and strategies.

The point is to test whether models can plan across time horizons — figure out the next keypress, the next sub-objective, and a multi-step strategy — without relying on flaky live gameplay. The benchmark ships with 5,000 offline multiple-choice questions and 20 stepwise online tasks broken into 62 subtasks, so evaluators can pinpoint exactly where a model breaks down.

The team tested 47 models spanning vision-language, GUI, coding, and game-agent families. GameHorizon is the first large-scale AAA-focused dataset that combines human-recorded actions, dense instructions, and reproducible offline evaluation in one place.

[16:29] NVIDIA Frames AI Agent Security as an Engineering Discipline

NVIDIA published a blog on September 21 making the case that securing AI agents is not a research problem, it's an engineering one. The post argues that defenses only count when they come with defined requirements, enforceable controls, a named owner, and documented evidence that the controls actually work.

The agent stack, the blog says, splits into three layers: models provide capabilities, harnesses organize context, tools, and workflows, and runtime environments provide the infrastructure where actions execute. Each layer carries its own security responsibilities, and protection has to hold across all three as data and instructions move through the system.

The concrete example the post walks through: an agent asked to update a customer record encounters malicious instructions in an attached document and tries to export the data to an outside destination. A network policy blocks the outbound transfer. Protected logs capture the attempted tool call, the authorization decision, and where it was headed. The agent holds permission to update the record but not to export it, and it cannot grant itself that extra access on its own.

NVIDIA points to its own open-source OpenShell runtime as one enforceable boundary — a sandboxed execution environment that enforces file, network, and process limits outside the agent's reasoning. Cisco's DefenseClaw adds governance on top, and JFrog integrates to scan and verify agent skills before they are allowed to run.

For evidence, the post calls for testing that blocks credential escalation attempts and unauthorized data transfers, repeated after any meaningful model or tool change, with a named owner signing off on readiness. It highlights CrowdStrike's SafeMind for repeated attack simulations, Palo Alto Networks' Prisma AIRS for continuous red teaming, Capital One's VulnHunter for AI-assisted code review, and ReversingLabs' Spectra Assure for catching malware in software dependencies.

The closing message is shared openly: post-incident findings should turn into repeatable tests, and the broader Open Secure AI Alliance exists to circulate what works across the security community.

[18:30] Thinking Machines Lab moves open-model inference to Crusoe Cloud under $65M deal

Thinking Machines Lab has signed a $65 million annual deal to run inference for its open models on Crusoe Cloud, the companies said on September 23. The arrangement routes production traffic through Crusoe Managed Inference, a service layer the company positions around throughput, price-performance, and reliability rather than raw training horsepower.

For builders, the practical question is what changes at the API. The model weights and behavior are unchanged — Thinking Machines Lab's open models stay open. What shifts is the serving stack underneath them, which is the piece that determines latency, uptime, and per-token cost. Crusoe is selling Managed Inference as a tuned environment rather than bare-metal access, so the lab hands off capacity planning in exchange for a managed agreement.

The deal is also a signal about where serious open-model serving is consolidating. Crusoe has pitched itself as a neocloud built for AI workloads, and landing a multi-year, eight-figure annual commitment from a notable AI lab is the kind of reference customer that shapes the next round of buyer evaluations. The announcement frames throughput and reliability as the headline wins, with price-performance as the third leg — a deliberate echo of how hyperscalers market their own inference tiers.

One thing worth watching is what this means for developers already running Thinking Machines Lab models elsewhere. If you're a current customer, expect the same model surface but a different provider behind the endpoints, which is worth re-benchmarking once traffic shifts. The larger question is whether Crusoe can convert this single anchor customer into a broader pattern of open-model labs outsourcing their inference rather than running it in-house.

[20:11] Jev returns probabilities instead of text, undercutting GPT-5 Nano on price

TypeSafe AI unveiled Jev earlier this month, and it's a different kind of model than most people are used to. The company calls these "System One" models; Simon Willison and Maggie Appleton both prefer the term "decision models." The shape is unusual: Jev still takes in text, but instead of generating text back, it returns floating point numbers — probabilities, confidence scores, and distributions over categories.

You build a "state" object describing whatever you're looking at — an article, a customer record, a chunk of semi-structured data — and then attach one or more typed questions. Jev answers each one with a structured probability instead of a sentence. There are three question shapes. "Noul" questions are yes/no — you state a proposition, and Jev returns a number between 0 and 1 representing how confident it is that the statement holds; the name, the company's CEO confirmed on Hacker News, is short for Bernoulli, as in the Bernoulli distribution. Choice questions take a list of labeled options and return a probability distribution across them. Score questions ask for a rating along a sequence of numeric levels.

The pricing is what makes it interesting for builders. Jev charges only for input tokens — output is free — and the input rate is $0.042 per million tokens. That undercuts OpenAI's GPT-5 Nano at $0.05 per million tokens, which TypeSafe positions as the benchmark it just beat. Because output is essentially free, Jev fits naturally into pipelines that fan out many parallel decisions per document.

For builders, that means Jev is worth a look anywhere you'd previously bolted a classification prompt onto a general LLM. It's a fit for routing, triage, or content moderation layers where you want a confidence number rather than a paragraph. Pair it with a regular text model downstream for anything that needs an actual reply.

[22:07] A Self-Hosted Browser That Lets AI Agents Skip Captchas

A new release of invisible_playwright_mcp gives AI agents a self-hosted way to browse the web without tripping common bot defenses. The project is a Python server that speaks Model Context Protocol, the same standard models use to call external tools, and it wraps Playwright, the popular browser-automation library, around an undetected, anti-detect stealth build of Firefox. The result is a browser session that mimics a regular user rather than a headless script, which helps it pass fingerprint checks and skip captchas. Version 0.70.0 shipped on September 23, and the GitHub repository now shows more than 31,600 stars.

For builders, the appeal is straightforward. Browser automation that used to require a paid captcha-solving service or fragile workarounds can now be handled inside a tool an agent can already call. Anything that needs a real session — logging in to a dashboard, scraping a site behind basic bot walls, filling multi-step forms, or driving a "computer use" loop — can run through this server in your own environment. That keeps credentials local and removes a third-party dependency from the loop.

The project is open-source and self-hosted, so the usual caveat applies: stealth browsers are arms-race tools, and an anti-detect Firefox from one project can be fingerprinted tomorrow. Watch upstream Firefox changes, pin the version, and keep usage within site terms and applicable law. If your agent is currently stuck behind a captcha wall, this is one of the more mature open options to try.

[23:38] OpenAI releases MentalHealthBench to test AI responses in mental-health conversations

OpenAI published MentalHealthBench on September 23. It is a benchmark — a standardized evaluation set — for measuring how AI systems respond in mental-health conversations, scored on two dimensions at once: being helpful and being safe.

What makes it distinct, according to OpenAI's announcement, is that the benchmark is expert-informed. Mental-health specialists were involved in shaping the realistic conversation scenarios the benchmark draws on. That detail matters because generic safety tests tend to miss the texture of an actual conversation — how a model handles someone in distress, talks through a panic moment, or knows when to point a user toward professional help.

The implied audience is developers and researchers building conversational AI for sensitive contexts — therapy-adjacent products, support bots, wellness apps, and the underlying models behind them. A public benchmark gives those teams a shared reference point instead of relying on hand-picked demos.

The framing also emphasizes a dual score. A model can be helpful but unsafe — engaging warmly while steering the user somewhere harmful — or safe but unhelpful, refusing or deflecting when the user genuinely needs guidance. MentalHealthBench tracks both, which means any single headline number will understate the picture. Teams shipping in this space now have a common yardstick, and they have to read both ends of it.

Watch next: whether third-party labs pick the benchmark up, and whether OpenAI releases baseline scores for its own models alongside it. A benchmark without published numbers is really just a rubric.

Chapters

  • 00:00 — Intro: Agent Stack Release Readout: Hermes Agent v2026.9.21 / Claude Opus 5.5 delivers Fable-class results at $4 in, $20 out / Head-to-head: GPT-6 Sol and GPT-6 Astra versus Anthropic's new flagship
  • 02:00 — Agent Stack Release Readout: Hermes Agent v2026.9.21
  • 03:10 — Claude Opus 5.5 delivers Fable-class results at $4 in, $20 out
  • 05:58 — Head-to-head: GPT-6 Sol and GPT-6 Astra versus Anthropic's new flagship
  • 07:23 — Grok 4.7 Lands: Bigger Model, Same $2/$6 Price
  • 09:09 — Nokia Open-Sources AnyJev: No-Training Layer Makes Open LLMs Reliable Decision-Makers
  • 11:40 — NVIDIA's Nemotron 3 Diarization Tracks Eight Overlapping Speakers Live
  • 13:21 — OpenAI extends Daybreak cyber access to Ukraine
  • 14:32 — Research digest: A benchmark for agent 'taste' on long tasks
  • 15:29 — Research digest: GameHorizon Puts AI Agents Through 5,000 Hours of AAA Gameplay
  • 16:29 — NVIDIA Frames AI Agent Security as an Engineering Discipline
  • 18:30 — Thinking Machines Lab moves open-model inference to Crusoe Cloud under $65M deal
  • 20:11 — Jev returns probabilities instead of text, undercutting GPT-5 Nano on price
  • 22:07 — A Self-Hosted Browser That Lets AI Agents Skip Captchas
  • 23:38 — OpenAI releases MentalHealthBench to test AI responses in mental-health conversations

Primary Links

Release Coverage Check

  • Hermes Agent — Latest stable verified: v2026.9.21, published 2026-09-21T18:10:55Z. Recent episode version tags detected: v2026.8.31, v2026.9.11, v2026.9.14, v2026.9.7. Selected missing version(s): v2026.9.21.
  • OpenAI Codex — Latest stable verified: rust-v0.156.1, published 2026-09-23T02:41:36Z. Recent episode version tags detected: rust-v0.153.0, rust-v0.153.2, rust-v0.155.0, rust-v0.155.1. No new stable release this cycle.
  • Claude Code CLI — Latest stable verified: 2.1.273, published 2026-09-15T18:06:34.098Z. Recent episode version tags detected: 2.1.236, 2.1.267, latest, stable. No new stable release this cycle.
  • Antigravity CLI — Continuous delivery model; no discrete release tags verified this cycle (latest build as of 2026-09-23). Recent episode version tags detected: @google/antigravity, v2.0.

Harness Version Reference

  • Hermes Agent — v2026.9.21
  • OpenAI Codex — rust-v0.156.1
  • Claude Code CLI — 2.1.273
  • Antigravity CLI — Continuous delivery (no tagged release verified this cycle)

🎙 Never miss an episode — subscribe now

SpotifyAppleiHeartYouTube MusicAmazonRSS
🎙 Subscribe to AgentStack Daily