Claude Opus 5.5 and GPT-6 Sol/Luna launched the same day . Opus 5.5: $4/$0.20/$20 per million tokens

Claude Opus 5.5 and GPT-6 Sol/Luna launched the same day (2026-09-22). Opus 5.5: $4/$0.20/$20 per million tokens (input/cached/output), 20% cheaper than Opus 5 with 60% cheaper cache reads, Anthropic framing it as 'the intelligence of Fable 5.1' at lower cost. GPT-6 Luna at $0.10/$0.50 is one of OpenAI's cheapest models ever; GPT-6 Sol at $2/$10. Anthropic also disclosed Opus 5.5 has 'extremely strong cyber capabilities' triggering new safeguards and a Cyber Verification Program -- but Simon Willison's own pelican-SVG test found Opus 5.5's maximum thinking level failing twice, $2.56 and 20 minutes per attempt, no output returned.

GPT-6 Astra decrypted a German Army Enigma message dated 10 July 1941 that had resisted solution since a crowd-cryptanalysis project began working it in 2005. A researcher supplied a known place-name crib from a related solved message; GPT-6 Astra built its own Enigma simulator and Bombe-equivalent search code and ran ~14.8 million key checks over ~10 hours. The cryptanalyst who verified the plaintext, Frode Weierud, says he is still analysing the logs to determine exactly how the model executed the break -- the result is corroborated, the autonomous-execution claim is not yet independently reconstructed.

Google open-sourced AX, a Kubernetes-native declarative orchestrator for 'billions of autonomous agent workloads' -- Task/Workspace/Gateway/Model primitives, checkpoint/resume, Apache 2.0. It debuted on GitHub Trending at +2,305 stars/day. This is the third distinct organization (after Cloudflare's computer/OS and Y Combinator's open-sourced QM) to independently ship the same architectural answer: cluster-scheduled, ops-layer infrastructure for agent fleets, rather than a new agent-building framework.

Foremerge, a Wednesday-rotation Show HN launch, detects when two coding agents working in isolated Git worktrees have silently undone each other's work -- a conflict of intent invisible to Git's own text-diff comparison. It is deliberately advisory-only (never locks a file, so one crashed agent can't stall the fleet) and deliberately keeps an LLM out of its own detection path (so identical inputs always produce identical findings) -- the same 'engineering code owns the steps that must not fail' argument this vault logged for Alibaba's open-code-review one week earlier, now applied to fleet coordination instead of code review.

Xiaomi open-sourced MiMo-V2.6-Pro and MiMo-V2.6-Flash under MIT license. Pro (1.02T total/42B active parameters, omnimodal, 1M context) scores 46 on Artificial Analysis' Intelligence Index -- tied for the top open-weight score in the world with Grok 4.7, ahead of DeepSeek V4.1 Flash (39) and Pro (36). Xiaomi disclosed its RL training cost: 30 RL steps over ~750,000 trajectories in under six days, for roughly $2.62M (Pro) and $0.85M (Flash) -- a rare vendor-published training-cost figure for a model landing at the top of an independent leaderboard.