Watch on YouTube · subscribe for the daily brief

Two labs shipped price cuts and called them models. The cheapest frontier token now costs a dime, and the most expensive one might not come from the model you paid for.

Two labs shipped price cuts and called them models. The cheapest frontier token now costs a dime, and the most expensive one might not come from the model you paid for.

Anthropic cut the price of its flagship by 20 percent on Tuesday and called it Claude Opus 5.5. About ninety minutes later OpenAI cut GPT-6 Sol and Luna in half and called that a launch too. Luna now costs ten cents per million input tokens.

Neither company led with a benchmark. Both led with a bill. That is the day.

What shipped

Opus 5.5 lists at $4 per million input tokens and $20 output, down from $5 and $25. Cache reads fell 60 percent to $0.20, and cache reads are most of an agent's spend. Anthropic says the model uses fewer tokens and finishes over 30 percent faster, so typical workloads run about 40 percent cheaper than Opus 5. It claims Fable 5.1 performance on most work, and it beats Fable on several agentic coding benchmarks.

GPT-6 Sol went from $4/$20 to $2/$10. Luna went from $0.20/$1.20 to $0.10/$0.50. OpenAI credits caching and inference improvements and says the prices are permanent, not promotional. Terra, the old cheap tier, is gone.

Yesterday I told you to budget per completed task. Today the labs did it for you. OpenAI's launch chart compares cost per task against Claude. Anthropic's launch talks about tokens per task. The unit of competition just changed.

Read the per-task line

Half the sticker price is not half the task price. On Artificial Analysis's same-harness test, Opus 5.5 at its default setting scored 51.2 for $1.34 per task. Sol's best run scored 47.5 for $1.06. Sol is half the list price and about 80 percent of the task price, and it scores lower. Sol's intelligence index score moved from 47 to 48. The Decoder calls the performance gain barely a needle move, and notes OpenAI's charts skip Terminal-Bench 4.0 and GDPval. Self-selected benchmarks, both sides.

The model ID is a suggestion

Here is the footnote The New Stack pulled out. Opus 5.5 ships with the safeguard tier Anthropic previously reserved for Fable, because it rates the model near Mythos 5.1 on biology and cybersecurity. Most cybersecurity requests are silently routed to Opus 4.8. Flagged biology work falls back to Opus 5. Artificial Analysis labels its Opus 5.5 result "Default Fallback." You call claude-opus-5-5 and you may get three different models across one agent run. If your eval assumes one model, it is measuring a mix. Thread 4 again: safety lives in access, and now access is a classifier you cannot see.

The open model that costs a dime

The day before, Xiaomi released MiMo-V2.6-Pro at $0.435 per million input tokens and $0.87 output. It scores 46 on the same index, the top open-weight result, and the same score Grok 4.7 posted at $2/$6. Weights are MIT licensed. The reinforcement learning run cost $2.62 million and Xiaomi streamed it live.

And the receipt. Anthropic's September 10 misuse report names Xiaomi in a distillation case, GTG-16008, tracking more than 400,000 exchanges over 20 days in which MiMo user sessions were allegedly routed through Claude. Xiaomi has not answered that in its technical report. The model is real and cheap. The provenance is disputed. Both things are true, and thread 9 keeps running.

Follow the money down the stack

Snorkel raised $350 million at a $3.5 billion valuation because training data demand is booming. Firecrawl, which scrapes the web for models, took $75 million. Heidi Health raised $340 million for agents inside hospitals. Baselayer raised $35 million for an agent trust layer. Known Systems spun out to make agents accountable to their owners. When tokens get cheap, the money moves to what feeds and governs the model.

Ben Lorica's essay lands on the same square. Your model might be fine, he argues, but your data is lying to it. Cheap inference does not fix a wrong semantic layer. It just makes the wrong answer cheaper.

Builder's bench

• Google open-sourced AX, a Kubernetes-style orchestrator for autonomous agents.
• JetBrains went all-in on agentic development and bet the IDE still matters.
• DuckDB now ships inside dbt v2. Transformers can run llama.cpp quants.
• Simon Willison shipped llm 0.36, llm-anthropic 0.29 and an alpha llm-typesafe for Jev decision models, and wrote the price war up.
• The New Stack: coding agents need a secrets-safe context boundary. Also its hands-on with Opus 5.5 finishing tasks, and Claude Cowork versus ChatGPT Work mode.
• Meta patched the Muse token exploit from yesterday and admitted Muse's resemblance to OpenClaw is not a coincidence.
• Teradata, Kantata and GitLab Duo all added governed-agent tooling. Rabbit's agent no longer needs an R1. Qualcomm shipped two AI phone chips.
• OpenAI: math advisory group, a claimed 100 open problems solved after a month of training, Astra cracking a 2005 Enigma message, and a repeat call for international standards. Self-reported except Enigma.
• Anthropic is building a biology lab where Claude directs robots. Toyota has workers training humanoids. AstroForge put AI in command of a spacecraft. Lab-grown neurons promise cheaper video.
• Security: Microsoft disrupted an AI-assisted platform that hit 12,000 targets; ClosedQuorum malware uses AI to pick its moves; Chinese actors hit WordPress and Zyxel.
• Policy and courts: British Columbia sued OpenAI over the Tumbler Ridge shooting and a separate suit wants a new school; Trump says the US is renaming AI "super intelligence"; a16z launched an academy with Palantir, Google and Meta.
• Reads: MIT Technology Review says don't be fooled by the summer of hype; "AI Has No Wisdom" on Hacker News; Starburst on banking use cases; Jun Kim joined Hugging Face for MLX.

Dropped: four TechCrunch event promos, Dyson's waterproofing problem, and MIT's border roundtable (noted Monday, still off-spine).

Through-line: the frontier is no longer priced by what a model can do. It is priced by what it costs to finish, and the vendor now decides which model finishes it.

One thing to do today: log the model ID that actually served each response in your agent traces, not the one you requested. Then rerun yesterday's per-task comparison with Opus 5.5 and Sol in it.

Sources covered

1 story · 65 sources (55 news, 4 company, 6 voices)

1. Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war
   🎙️Simon Willison↗📰SiliconANGLE↗📰SiliconANGLE↗📰The Verge↗📰TechCrunch↗📰SiliconANGLE↗🎙️Simon Willison↗📰SiliconANGLE↗🎙️Simon Willison↗📰TechCrunch↗📰Ars Technica↗🎙️Simon Willison↗📰The Verge↗📰BleepingComputer↗📰The Decoder↗📰TechCrunch↗🎙️Simon Willison↗📰The New Stack↗📰Ars Technica↗📰The New Stack↗📰Ars Technica↗📰TechCrunch↗📰The New Stack↗📰BleepingComputer↗📰Hacker News↗📰The New Stack↗📰TechCrunch↗📰The New Stack↗📰The Decoder↗📰Ars Technica↗🎙️Ben Lorica↗📰The Verge↗📰The New Stack↗📰TechCrunch↗📰The Verge↗📰Hacker News↗📰The Verge↗📰The Decoder↗📰TechCrunch↗📰TechCrunch↗📰SiliconANGLE↗📰TechCrunch↗📰InfoQ↗📰TechCrunch↗📰The Decoder↗📰Hacker News↗📰MIT Technology Review↗📰Ars Technica↗📰SiliconANGLE↗📰SiliconANGLE↗📰SiliconANGLE↗📰The New Stack↗📰Hacker News↗📰SiliconANGLE↗🏢Starburst↗📰InfoQ↗📰The Verge↗📰The Decoder↗📰The Decoder↗📰MIT Technology Review↗📰The Decoder↗📰Al Jazeera↗🏢DuckDB↗🏢Hugging Face↗🏢Hugging Face↗