Anthropic cuts Opus prices 40% as OpenAI halves GPT-6 Sol and Luna
Anthropic released Claude Opus 5.5 on 22 September, and OpenAI shipped GPT-6 Sol and GPT-6 Luna the same day. Neither release leads on new capability, though both lead on price.
Anthropic says Opus 5.5 costs 40% less to run than Opus 5 at default settings on typical workloads. Input and output tokens are $4 and $20 per million, 20% below Opus 5. Cache reads, which the company says make up the majority of agentic and coding work costs, drop to $0.20 per million, a 60% cut. Anthropic also claims that output generates more than 30% faster.
OpenAI went further on the headline number. Its own API pricing table lists GPT-6 Sol at $2 per million input tokens and $10 per million output, exactly half the $4 and $20 charged for GPT-5.6 Sol. Luna sits at $0.10 and $0.50, against $0.20 and $1.20 for GPT-5.6 Luna. CNBC reported the move as a 50% cut against GPT-5.6 promotional pricing.
| Model | Input, per 1M tokens | Output, per 1M tokens |
|---|---|---|
| Claude Opus 5.5 | $4 | $20 |
| Claude Opus 5 | $5 | $25 |
| GPT-6 Sol | $2 | $10 |
| GPT-5.6 Sol | $4 | $20 |
| GPT-6 Luna | $0.10 | $0.50 |
| GPT-5.6 Luna | $0.20 | $1.20 |
One of the things we’re continuing to innovate on is how to make that thinking, how to make the answering more efficient, so it uses less tokens depending on your effort setting.
Dianne Penn, head of product management, research and labs at Anthropic, via CNBC
That matters because per-token price isn’t the whole bill. Anthropic’s 40% figure combines a cheaper token with fewer tokens spent per task, which is a different claim from the 20% sticker cut. If you’ve ever worked through what an AI feature actually costs to run, the cache read line is the one to watch here.
The benchmarks are close, and Anthropic says so itself
On Anthropic’s own numbers, Opus 5.5 scores 66.4% on Terminal-Bench 4.0 against 57.9% for GPT-6 Astra and 52.3% for Opus 5. On FrontierCode v1.1 (Main) it’s 54.4% to Astra’s 53.3%. The company adds its own caveat, that at these capability levels benchmark margins have become a less reliable guide to real-world differences. Ars Technica read the gains as real but modest, and framed Anthropic as playing catch-up after Astra.
The timing is the awkward part. Both releases are the first since Anthropic CEO Dario Amodei called for pacing the frontier. CNBC reports that former Anthropic researcher Jacob Coxon set the debate off on 8 September, posting to X that he had quit and warning both labs were “gambling with our lives”. Sam Altman and Elon Musk joined Amodei’s call.
Anthropic’s answer is that Opus 5.5 is its best-behaved model so far. It says the model posted the strongest scores to date on an automated behavioural audit covering nearly 2,000 scenarios, and that outside evaluators including METR and Frontier Design tested it before release.
Opus 5.5 attempted to circumvent boundaries around 85% less often than Opus 5 or Claude Mythos 5.1, and every attempt it made was low severity and self-reported.
Anthropic, Introducing Claude Opus 5.5
There’s a cost to that. Opus 5.5 ships with cybersecurity and biology safeguards similar to those on Fable 5.1, and Anthropic says most cybersecurity tasks get re-routed to Opus 4.8. AWS puts it plainly in its launch post: requests are refused more often than on previous Opus versions. So the cheaper model is also the more restricted one, which is a trade some teams won’t want.
Availability is broad on both sides, so the cut reaches most buyers at once. Opus 5.5 runs on Amazon Bedrock, Google Cloud and Microsoft Azure, and on the Claude Platform as claude-opus-5-5, with Sonnet 5.5 and Haiku 5.5 due in the coming weeks. Engadget notes that GPT-6 Sol and Luna reach ChatGPT Work and Codex for Plus, Pro, Business and Enterprise customers, while free users and Go subscribers get Luna only, in the desktop app.
CNBC frames the squeeze as competition from cheaper open-weight models, naming Alibaba, Moonshot AI and DeepSeek. That pressure is why open weights keep reshaping the price sheet, and it’s also why the interesting question isn’t which model tops a leaderboard. It’s whether a cheaper flagship pulls work back from the routers, or whether teams keep asking how often the entry-tier model is enough. The next data point is Sonnet 5.5.
Get the daily rundown
One email each weekday with the AI news that matters, every claim linked to its primary source.
Free, one email each weekday, unsubscribe in one click. We never sell or share your address.
