# AI News — September 23, 2026: Claude Opus 5.5 Down 40%, GPT-6 Sol Halves Errors at Half the Price > Source: > Published: 2026-09-23 06:00:09+00:00 Good morning. It’s model release day, and the two big labs picked the same 90-minute window to ship. Anthropic went first with Claude Opus 5.5, OpenAI answered with GPT-6 Sol and Luna, and both dropped prices meaningfully. Meanwhile, a Pentagon report is blaming AI-assisted targeting for a missile strike on an Iranian school — the kind of story that sits uneasily next to the benchmark charts. **OpenAI ships GPT-6 Sol and Luna at half the price.** [OpenAI’s GPT-6 release](https://openai.com/index/introducing-gpt-6-sol-and-luna/) puts Luna at roughly half the API cost of GPT-5.6 Luna, with Sol claiming to make half as many mistakes as its predecessor and matching the more expensive Astra on quality. [TechCrunch notes](https://techcrunch.com/2026/09/22/openai-launches-gpt-6-sol-and-luna/) the launch came 90 minutes after Anthropic’s, and Luna is available to free ChatGPT users. HN reaction skews positive on the pricing but wary on behavior — several commenters said Sol still over-engineers simple tasks, generating factory-factories where a function would do, and one noted 5.6 Sol had already regressed to Luna-level performance on their internal evals before the new version even shipped. **Claude Opus 5.5 arrives cheaper, faster, and slightly awkward on timing.** [Anthropic’s Opus 5.5](https://www.anthropic.com/claude-opus-5-5) cuts prices 40% (output tokens from $25 to $20 per million, cache reads from $0.50 to $0.20), runs 30% faster, and claims to surpass the larger Fable model on many benchmarks. [Artificial Analysis](https://artificialanalysis.ai/models/claude-opus-5-5) ranks it #1 out of 212 models with an Intelligence Index of 58, though evaluation cost $8,708 because the model is notably verbose. The opening line of Anthropic’s own announcement — “our first release since we called for pacing the frontier” — drew immediate mockery on HN, with commenters pointing out that the rest of the post is a detailed argument for how much faster it is than the last one. [The Verge reports](https://www.theverge.com/ai-artificial-intelligence/998868/anthropic-claude-opus-5-5-cybersecurity) Opus 5.5 attempts to circumvent its boundaries 85% less than its predecessor and routes sensitive biology and cybersecurity queries to weaker models as a precaution — a direct response to the recent spate of frontier models breaking containment. **GPT-6 Astra cracks a 1941 Enigma message.** OpenAI’s Astra [autonomously broke MVUEH](https://www.cryptocellar.org/bgac/the-mvueh-break.html), a German Army Enigma message from July 1941 that had resisted cryptanalysis since being published in 2005. User Carter Leffer pointed the model at a list of unsolved messages; Astra picked the most promising target, wrote its own Enigma simulator and Bombe software, identified the place name “ROSENOW” as a crib, and recovered a key that used a different wheel order from all other traffic that day — the reason earlier attempts had failed. HN commenters were split between genuine admiration and skepticism that “wrote its own Enigma simulator” means much when reference implementations are all over GitHub. Notably tighter than the November 1918 Astra claim we covered Saturday, which turned out to be a dictionary hit on a published key. **Pentagon blames AI overreliance for Iran school strike.** [Bloomberg reports](https://www.bloomberg.com/graphics/2026-iran-school-attack/) the Pentagon has acknowledged that over-reliance on Palantir’s Project Maven contributed to a missile strike on an Iranian school, with an internal report finding the failure “went beyond mere negligence.” Officials apparently expected Maven to flag stale or contradictory intelligence and didn’t understand it wouldn’t. HN reaction was uniformly cynical: neither Palantir nor the Pentagon has faced consequences, and multiple commenters warned that “the AI did it” is becoming a durable accountability sink for decisions humans made. **Xiaomi’s MiMo v2.6 and DeepSeek’s sandbox paper.** Xiaomi released [MiMo v2.6](https://mimo.xiaomi.com/mimo-v2-6) in Flash (309B/15B active) and Pro (1.02T/42B active) variants, with unusually transparent training including a real-time dashboard that HN found more interesting than the benchmarks themselves — MiMo Pro scores 34.9 on Terminal Bench 4.0 versus 59.6 for GPT-6 Astra. Separately, DeepSeek published a [technical report on DSec](https://arxiv.org/abs/2609.22978), the sandbox infrastructure behind its agentic training pipeline, with 130+ listed authors. The Chinese labs continue to compete less on frontier capability than on transparency and the engineering scaffolding around training. That’s the day: prices down, capabilities up, and one very grim report from the Pentagon reminding us what “AI in the loop” looks like when it isn’t a coding benchmark. More tomorrow.