The biggest model launch day of the quarter, plus a funding round, a security disclosure that deserves more attention than it got, and a lawsuit.
Anthropic Ships Claude Opus 5.5
Released 22 September at $4 per million input tokens and $20 per million output, a 20% list cut from Opus 5. Cache reads dropped 60% to $0.20 per million. Context window up to 1M tokens, 128K max output, June 2026 knowledge cutoff.
It takes Terminal-Bench 4.0 at 66.4% against GPT-6 Astra's 57.9%, and GDPval-AA v2.1 at 1846 Elo. It loses AutomationBench to Astra 40.0% to 41.4% and Terminal-Bench-Science 58.7% to 64.6%.
Anthropic claims 40% lower running cost on typical workloads. The list price only moved 20% - the rest depends on cache reuse and 30% faster output generation, so it is a claim about your pipeline rather than a number you can put in a budget.
Four API changes break Opus 5 code: thinking cannot be disabled, tool_choice: "any" returns a 400, thinking blocks are bound to the model and conversation so replays fail, and the computer_20251124 tool is rejected. Live on the Claude API, Bedrock, Google Cloud and Microsoft Foundry. Sonnet 5.5 and Haiku 5.5 follow in the coming weeks.
Full breakdown of Opus 5.5 pricing, benchmarks and the four breaking changes
OpenAI Counters With GPT-6 Sol and Luna
Same day. GPT-6 Sol at $2 in / $10 out, GPT-6 Luna at $0.10 in / $0.50 out - both half the price of their GPT-5.6 equivalents. Cached reads carry a 90% discount. Flagship Astra is unchanged at $10 / $50.
Luna is the story. It scores 66.6% on DeepSWE v1.1 against Sol's 68.8% and Astra's 74.1%, at one twentieth of Sol's price. But it drops to 20.7% on AutomationBench against Sol's 33.2%, and carries a 7.6% factual error rate against Astra's 3.9%. It writes code well and plans badly.
Available as gpt-6-sol and gpt-6-luna in the API, and in ChatGPT for Plus, Pro, Business, Enterprise and Edu through ChatGPT Work and Codex. Free and Go users get Luna in the desktop app.
Full GPT-6 Sol and Luna pricing and benchmark breakdown
What Two Launches in One Day Means
Three frontier price cuts landed inside 48 hours: Grok 4.7 on 21 September at $2 in / $6 out, then Opus 5.5 and the GPT-6 pair on the 22nd. Nobody led with a capability claim. Everyone led with price.
If you are running anything on a model released before this week, re-run your cost model. The floor moved.
Snorkel AI Raises $350M at a $3.5B Valuation
Led by Insight Partners and S32, roughly tripling the company's previous valuation. Snorkel sells training-data infrastructure for frontier model builders - the unglamorous layer underneath every launch above.
The round is a reasonable proxy for where the money is going. As model prices fall, the differentiator moves from architecture to data, and the companies selling data tooling get repriced accordingly.
Cisco Talos Discloses an Autonomous AI Implant
Talos released a toolkit called CAIRN alongside a disclosure of CLOSEDQUORUM, which it describes as the first reported fully autonomous multi-model AI command-and-control implant.
The phrase that matters is multi-model. An implant that can route between several providers is not stopped by any one vendor cutting it off. This is worth more attention than it is getting on a day dominated by launch news.
Palo Alto Networks Launches Unit 42 Frontier AI Defense
Announced 22 September: Unit 42 Continuous Frontier AI Defense, using several models together - Anthropic's Claude Mythos 5, OpenAI's GPT-5.6-Cyber and open-weight models - for vulnerability detection.
Both sides of the security line are now running multi-model architectures for the same reason: no single model is reliable enough alone, and ensembles catch what one misses.
Z.ai Open-Sources ZCode
Following the disclosure that ZCode was silently uploading workspace data overseas, Z.ai has released it under Apache-2.0.
Open-sourcing after an exfiltration finding is the right move and also the cheapest available one. It makes the behaviour auditable going forward. It does nothing about the 313MB that already left, and if you ran ZCode you should still rotate anything that was in those workspaces.
What ZCode uploaded and how to check whether you were affected
Also This Week
- British Columbia sued OpenAI and Sam Altman over the February 2026 Tumbler Ridge school shooting. Filed 21 September.
- Meta Muse for Mac zero-day. Security researcher Patrick Wardle disclosed a flaw letting local applications hijack Muse authentication tokens on macOS.
- Grok 4.7 shipped 21 September at $2 in / $6 out per million tokens, 71% on DeepSWE v1.1. Our benchmark breakdown.
What To Do About It This Week
- Re-run your cost model. Three price cuts in 48 hours. Whatever you budgeted in August is wrong.
- Test Luna on your highest-volume task. $0.10 per million input is a different category. If it holds on your workload, nothing else this week matters as much.
- Audit for the Opus 5.5 breaking changes before you change any model string - particularly if you store and replay assistant turns.
- If you ran ZCode, rotate your credentials. The Apache-2.0 release does not undo the uploads.
Sources