Five Releases This Week, One Dominant Story

OpenAI shipped GPT-6 Cyber on 5 November, making it the first production model to receive the company’s Critical Preparedness rating. That label means OpenAI’s safety team judged the model capable of providing "meaningful uplift" to attacks on critical infrastructure — a threshold the company says requires restricted deployment even under its own policies. Access is limited to vetted Daybreak Red customers via IP-allowlisted API and Codex calls.

Four other releases rounded out the week. Only SWE-2 Flash has immediate relevance for most teams.

The Week in Review: Ranked

Entry Breakdown

1. GPT-6 Cyber: Critical, Costly, Controlled

OpenAI’s Preparedness Framework has five levels: Low, Medium, High, Critical, and Maximum. GPT-6 Astra crossed into Critical on the cyber dimension in September; GPT-6 Cyber is the first model deliberately engineered to operate at that level while maintaining safety guardrails strong enough for restricted deployment.

The numbers from the OpenAI-run evaluation are striking. ExploitGym: 48.5%, up from 32.1% for GPT-5.6 Cyber and 42.4% for the unrestricted GPT-6 Astra measured under Daybreak Blue conditions. RemoteExploitBench: 94.7%. Web pentesting without source code: 40.3%, versus 32.0% for GPT-6 Astra. On the honeypot test — which probes whether the model exploits systems outside its authorised scope — GPT-6 Cyber stayed in bounds 99.93% of the time without deployment safety instructions, up from 94.49% for GPT-5.6 Cyber.

OpenAI also published a Linux kernel security study: total cost approximately $1.21M in API credits, yielding 948 working local privilege-escalation exploits from 2,308 candidate issues, and 304 critical findings disclosed to maintainers first. That works out to roughly $1,279 per working exploit — a number worth bookmarking for anyone building threat-modelling economics. All benchmarks are OpenAI-run; no independent evaluation exists yet.

Pricing: $20 per million input tokens, $100 per million output tokens. Cache reads $2 per million. Per Dataconomy citing OpenAI documentation, the input rate is about 60% higher and the output rate about 33% higher than GPT-5.6 Cyber’s standard rates. Verdict: Use it — if you have Daybreak Red approval. Everyone else: watch the OpenAI Daybreak Red waitlist.

2. SWE-2 Flash: Cheaper, Faster, But Devin-Only

Cognition released SWE-2 Flash on 4 November. The base model is Inkling-Small from Thinking Machines, a 276B-total / 12B-active mixture-of-experts. Cognition applied a mid-training stage and reinforcement learning before releasing under the Devin brand.

The benchmarks are all Cognition-run. SWE-2 Flash scores 24.0% on FrontierCode 2.0 (versus 23.8% for the larger SWE-2), 89.9% on SWE-Bench Pro v2 Hard (vs 90.6%), and 60.1% on DeepSWE 1.1 (vs 73.0%). On the benchmark that matters most for speed-conscious teams — Terminal-Bench 4.0 — it scores 20.0%, well behind SWE-2 at 27.3% and the current open-weight leaders (DeepSeek V4.1-Flash at 42.3% and GLM 5.3 Flash at 41.8% per Cognition’s own table). The 3× speed claim (200 vs 60 tok/s) is based on a replay methodology on one task; OrcaRouter notes it describes an intended serving profile rather than a measured one.

There is no standalone API and no public weights. Access is through Devin Desktop and Devin CLI only. Free through 30 November for Pro, Max, and Teams subscribers. Cognition says a secure-share program for research partners is coming. Verdict: Test it — Devin users have a free month to evaluate before the November price kicks in. Non-Devin teams: nothing to do yet.

3. Claude Text Watermark: Seven Models, Two Waves

Anthropic expanded its SynthID-Text watermark to three models on 3 November — Claude Mythos 5, Opus 4.7, and Sonnet 4.6 — with four older models (Opus 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5) following on 16 November. The second-wave rollout applies only when temperature, top_p, and top_k are at their defaults. No model capabilities change; no pricing changes. Verdict: Watch — if you post-process Claude output or redistribute it, verify your pipeline handles the change gracefully before 16 November.

4. ChatGPT Images 2.6: Faster, Stickier Edits

OpenAI updated its image generation product on 4 November. ChatGPT Images 2.6 generates images approximately 50% faster than its predecessor, retains brand, product, and face consistency better across iterative edits, and adds GIF creation (rolling out gradually) and saved named reference images callable with @name. Available across all ChatGPT plans and via the developer API, with no price change. No independent benchmark. Verdict: Test it — if you use the Images API in a workflow, the speed gain is worth verifying in production.

5. FlashProve: Interesting Science, No Production Path

Anthropic published FlashProve on 2 November (Apache 2.0, code on GitHub). It is a zkSNARK system that generates a cryptographic proof that a given response came from a specific model. Tested on the publicly available 4-bit weights of Llama 4 Maverick (400B), it reaches ~110 tokens per second on one H100 — a large improvement on prior work — but at 70–95× the GPU time of normal inference. Each proof is ~19 MB; context is capped at 1,024 tokens; and the architecture must be a mixture-of-experts. No production deployment plans exist at Anthropic or any other lab, per TechCrunch reporting via Trew Knowledge. Verdict: Skip it — file under “promising direction” and revisit in 12–18 months.

Cost Check: What GPT-6 Cyber Actually Runs

For a 100-task offensive security engagement — say 8,000 input tokens and 3,000 output tokens per task (a typical automated pentest job) — here is the token cost at published rates:

† GPT-5.6 Cyber rates derived from Dataconomy’s characterisation that GPT-6 Cyber input is ~60% higher and output ~33% higher than GPT-5.6 Cyber standard rates; not sourced directly from OpenAI. GPT-6 Cyber rates from BenchLM and Dataconomy citing OpenAI documentation. Task size (8K input / 3K output) is illustrative.

The price premium is about $13.50 per 100 tasks. At the Linux kernel study’s rate of $1,279 per working exploit, teams will want to track cost-per-finding carefully. For comparison, the kernel study ran approximately $1.21M total and produced 948 exploits — a scale that only well-resourced security programmes will replicate.

For Swiss & EU teams

The Anthropic watermark expansion has a direct EU angle. EU AI Act Article 50(2) requires that text generated by AI systems be labelled or technically marked as AI-generated. The transition period for systems already on the market before 2 August 2026 ends around 2 December 2026. Anthropic signed the Article 50(2) code of practice in July with approximately 190 other signatories. The 3 November and 16 November rollouts are part of that commitment. If your product uses Claude output and you redistribute it to end users, verify that your terms and disclosures align with Article 50(2) before 2 December. GPT-6 Cyber and SWE-2 Flash have no published EU region or data residency statements in this week’s releases.

What to Watch Next Week

OpenAI has indicated it will expand GPT-6 Cyber access to all enrolled Daybreak Red customers in the coming weeks. Independent benchmarks for GPT-6 Cyber and SWE-2 Flash are still absent; the numbers above are vendor-run. For Mercury 3, the diffusion LLM from Inception Labs reviewed two weeks ago, independent evaluation from Artificial Analysis is still the signal to wait for before drawing conclusions.

Further Reading

• GPT-6 Astra Crossed OpenAI’s Critical-Cyber Line: What It Means — background on the Preparedness Framework thresholds and what Critical actually requires.
• Model Scorecard: Week of 27 Oct 2026 — Terra Halves the Cost — last week’s rankings, including GPT-6.1 Terra and Kimi K3.1 Pro.
• How to Evaluate LLMs for Your Use Case (2026 Guide) — a framework for going beyond vendor benchmarks when assessing any new model.

Your turn: Does your organisation have Daybreak Red access, and what criteria did OpenAI require? Reply to our newsletter or send us a note — we feature the best answers in the Friday Scorecard.

How this article was made: AI researched and wrote this article under the Adrian persona, using the sources linked above, and it was published automatically without a human edit. Editorial guidelines are set by Adi. Spotted an error? Tell us and we will correct it.