Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Sol and GPT-6 Luna launched on the same afternoon, and both led with the invoice rather than the benchmark. Opus 5.5 claims to do Fable-level work for 40 per cent less than Opus 5, while GPT-6 Sol and Luna bring Astra-era gains to cheaper tiers at half the API price of the models they replace. The quieter, more useful half of both launches is caching: cheaper cache reads from Anthropic and a smarter caching system from OpenAI are where agent builders will actually feel the saving. Around the headline pair, Alibaba confirmed Qwen 4 is in training and unveiled its own AI accelerator, Microsoft prepared steep Copilot discounts, and Xiaomi's MIT-licensed MiMo kept undercutting everyone on open weights. Nobody moved the top of the capability chart today, and that is the point: the fight has settled firmly onto price.
Claude Opus 5.5 vs GPT-6 Sol and Luna: API pricing at a glance
| Model | Input, per 1M tokens | Output, per 1M tokens | Change |
|---|---|---|---|
| Claude Opus 5.5 | $4.00 | $20.00 | 40 per cent cheaper to run than Opus 5; cache reads $0.20 |
| GPT-6 Sol | $2.00 | $10.00 | Half the price of GPT-5.6 equivalent |
| GPT-6 Luna | $0.10 | $0.50 | Half the price of GPT-5.6 equivalent |
| Xiaomi MiMo-V2.6 Pro | $0.435 | $0.87 | New, open weights (MIT) |
| Xiaomi MiMo-V2.6 Flash | $0.14 | $0.28 | New, open weights (MIT) |
Claude Opus 5.5 pricing and limits: Fable-level work at 40 per cent less than Opus 5
If you pay for Claude by the token, or run long agentic sessions in Claude Code, this is the release to test this week. Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work, and on typical workloads costs 40 per cent less to run than Opus 5, because it is cheaper per token and uses fewer tokens per task. List prices drop to 4 dollars input and 20 dollars output per million tokens, and cache reads fall 60 per cent to 0.20 dollars per million, which is where most coding and agent costs sit.
Subscribers also get higher five-hour usage limits across Pro, Max, Team and seat-based Enterprise plans, plus a saveable rate limit reset. Anthropic is making a point of writing quality this time, saying the model puts key information first and follows your style rules more closely. It extends the Fable 5.1 price cut from earlier this month down the range, though most cybersecurity tasks are rerouted to Opus 4.8 by default, and heavy biology work needs the new Life Sciences Verification Program.
GPT-6 Sol and GPT-6 Luna: Astra-era gains at half the API price
For anyone building on the OpenAI API, the headline is price: GPT-6 Sol is 2 dollars input and 10 dollars output per million tokens, and GPT-6 Luna is 0.10 and 0.50, both 50 per cent cheaper than the GPT-5.6 versions they replace. OpenAI says both models were trained with similar methods to GPT-6 Astra, with Sol making roughly half as many factual mistakes as its predecessor on its internal evaluation. In ChatGPT Work and Codex they roll out gradually from today to Plus, Pro, Business, Enterprise and Edu users, and Free and Go users can use Luna in the desktop app.
The model IDs are gpt-6-sol and gpt-6-luna. Astra stays the top model, so the practical question is whether Sol is now good enough for the everyday work you currently route to something more expensive.
GPT-6 prompt caching: higher hit rates, a new dashboard and cache-friendly controls
This one matters if you run persistent agents or long conversations, because cached input is billed at up to 90 per cent off. OpenAI says the GPT-6 family now gets higher cache hit rates by default, with discounts applied to shared prefixes reused within a 30-minute window. You can now change reasoning effort or switch tools on and off mid-session without breaking the cache, and explicit breakpoints let you choose where cached prefixes end.
A new Prompt Caching Dashboard and diagnostics tool show how much input is cached and why a request missed. For agent builders, it is the kind of plumbing change that can quietly cut your bill more than a price drop.
Alibaba confirms Qwen 4 is in training and unveils the Zhenwu V900 AI chip
No new model you can use today, but a clear signal of what is coming if you build on open Qwen weights. At its Apsara Conference in Hangzhou, Alibaba said Qwen 4 is now in training on a new architecture, with Qwen 4.5 and Qwen 5 planned at 5 to 10 trillion parameters. It also announced Qwen3.8-LiveTranslate for simultaneous interpretation, new Qwen Audio models, Qwen-Image 3.1 later this year, and Qwen Intelligence, an agent platform for smartphone makers.
On hardware, the T-Head Zhenwu V900 accelerator claims three times the performance of the previous chip, with commercial availability planned for early 2027. Alibaba's claims about self-improving Qwen3.8-Max runs are its own figures and have not been independently checked.
Microsoft Copilot discounts of 30 to 50 per cent for bulk enterprise buyers from October
If your organisation has been holding off on Copilot licences because of cost, the maths is about to change. Microsoft will offer enterprise customers buying in bulk discounts of between 30 and 50 per cent on Copilot subscriptions, taking effect in October. The timing lines up with an upgraded Copilot app that pulls more of Microsoft's AI tools into one place.
It reads as a push to win seats before rivals' cheaper models make the per-user price harder to justify, exactly the pressure the day's model launches are applying. Note that this report comes from a financial news service rather than a Microsoft announcement page, so check the terms with your Microsoft account rep.
Xiaomi MiMo-V2.6 Pro and Flash: trillion-parameter open weights under an MIT licence
This is the most interesting open-weight release of the day, especially if you want a model that handles text, images, video and audio together without stitching tools. Xiaomi's MiMo-V2.6-Pro is a 1.02 trillion parameter mixture-of-experts model with a 1 million token context window, and the weights are on Hugging Face under an MIT licence. Xiaomi claims it scores 46.32 on the Artificial Analysis Intelligence Index, the top open-weight score in its comparison, though it still trails Claude Fable 5.1 and GPT-6 Astra overall.
API pricing is very low: Flash is 0.14 dollars input and 0.28 dollars output per million tokens, and Pro is 0.435 and 0.87, with both also on OpenRouter. It confirms the open-model crown we flagged yesterday, and if you are cost-sensitive and comfortable self-hosting or using a Chinese provider, it is worth benchmarking against your current setup.
Industry themes
Price is now the main battleground. Anthropic and OpenAI both launched on the same afternoon, and both led with cost: Opus 5.5 at 40 per cent less than Opus 5, and GPT-6 Sol and Luna at half the price of their predecessors. When two rivals ship on the same day and neither claims a new capability record, the market has decided the contest is about dollars per useful task.
Caching is where the real savings sit for agent builders, with cheaper cache reads from Anthropic and a smarter caching system from OpenAI. Before you compare headline prices, check your own cache hit rate, because a mid-session tool switch that no longer breaks the cache can cut a bill more than a list-price drop.
Open weights are also closing in, with Xiaomi's MiMo-V2.6 undercutting everyone on price even though it still trails the top closed models, and Alibaba signalling Qwen 4 and its own silicon are on the way. Both big launches were posted on the companies' own X accounts at the same time as their blogs, so the social scan confirmed them first-hand rather than surfacing anything the press missed.