On Tuesday, 22 September 2026, Anthropic released Claude Opus 5.5. According to Decrypt and TechCrunch, OpenAI announced GPT-6 Sol and GPT-6 Luna roughly ninety minutes later. The two launches were built very differently. Their pitch was not: frontier-grade capability for noticeably less money. This is the first week in which the two leading Western labs have competed openly on price per task as much as on benchmark scores. For anyone running coding agents or document pipelines in production, that matters more than any leaderboard.
What was released
Anthropic — Claude Opus 5.5. This is the first model in Anthropic’s 5.5 generation. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow “in the coming weeks”. Opus 5.5 is available through Anthropic’s own platform and on AWS, Google Cloud and Microsoft Azure. The headline claim is that it performs at the level of Claude Fable 5.1, Anthropic’s larger and more expensive tier, on most work, while costing about 40% less than Opus 5 on typical workloads. Anthropic also says it generates output more than 30% faster than Opus 5.
OpenAI — GPT-6 Sol and GPT-6 Luna. These are the mid and small tiers under GPT-6 Astra, the flagship OpenAI released on 3 September. Sol is meant for complex work such as coding. Luna is meant for fast, high-volume clerical tasks: summarising, extraction and quick answers. Both went live in the API and in Codex. In ChatGPT they reached Plus, Pro, Business, Enterprise and Edu users, while Free and Go users get Luna in the desktop app only. OpenAI’s main quality claim is that Sol makes “about half as many mistakes” as GPT-5.6 Sol.
The numbers
List prices per million tokens, as published and cross-checked across outlets:
| Model | Input | Output | Previous generation |
|---|---|---|---|
| Claude Opus 5.5 | $4.00 | $20.00 | Opus 5: $5 / $25 |
| GPT-6 Sol | $2.00 | $10.00 | GPT-5.6 Sol: $4 / $20 |
| GPT-6 Luna | $0.10 | $0.50 | GPT-5.6 Luna: $0.20 / $1.20 |
| GPT-6 Astra (reference) | $10.00 | $50.00 | — |
Caching is where the less obvious savings are. Anthropic cut cache reads on Opus 5.5 by 60%, from $0.50 to $0.20 per million tokens, and cache writes drop from $6.25 to $5. OpenAI says eligible cached input now gets up to a 90% discount. For agent loops that resend the same long system prompt, tool schemas and repository context on every turn, the cache-read price often matters more than the headline input price. Anthropic also offers a “fast mode” for Opus 5.5 at $8 / $40, which it says runs up to 2.5 times faster.
Each lab also published benchmark figures, and they mostly cover different tests:
- Anthropic reports Opus 5.5 at 66.4% on Terminal-Bench 4.0 (Fable 5.1: 55.8%, Opus 5: 52.3%, GPT-6 Astra: 57.9%). It reports 54.4% on FrontierCode v1.1 (Astra: 53.3%), 81.8% on OSWorld 2.0 and 67.7% on Humanity’s Last Exam (Astra: 57.2%).
- OpenAI reports that Sol scores 33.2% on AutomationBench, a business-workflow test across 47 tools, at $0.27 per task. It puts Claude Opus 5 at 26.9% on the same test, at about 11 times the cost. OpenAI also cites 60.5% on OSWorld 2.0 and 56.4% on Agents’ Last Exam. For Luna, MarkTechPost reports 66.6% on DeepSWE v1.1.
Read the benchmarks carefully
Neither lab’s charts include the other’s new model. Anthropic compared Opus 5.5 with GPT-6 Astra, OpenAI’s much more expensive flagship. OpenAI compared Sol with Claude Opus 5, the model Opus 5.5 replaced about ninety minutes before Sol launched. So there is no head-to-head comparison of the two new models from either company.
The overlapping benchmark shows the problem. Anthropic lists Opus 5 at 74.0% on OSWorld 2.0. OpenAI’s comparison, as reported by Decrypt, puts Opus 5 at 60.3% on the same named benchmark. MarkTechPost describes OpenAI’s figure as an “offline” score. The gap is almost certainly down to different harnesses, settings or subsets, not to one company being wrong. But it means you cannot take Sol’s 60.5% from one chart and Opus 5.5’s 81.8% from another and compare them. Vendor benchmarks run with vendor harnesses tell you about progress within one model family. They say much less about comparisons across families.
The “cost per task” figures that both labs now emphasise deserve the same caution. Cost per task depends on the reasoning-effort setting, on how many tokens a model spends thinking, and on cache hit rates. None of those carries over from a lab’s evaluation to your workload. A model with cheaper tokens can still cost more per task if it takes more turns to finish.
Why this matters for the future of IT
Three points stand out.
Price cuts are coming from efficiency, not only from discounts. Anthropic says Opus 5.5 needs less compute and runs faster. OpenAI credits inference and caching improvements. If the gains are real, the cost of work that needed a top-tier model six months ago is falling by roughly half per generation. That changes what is economical to automate: large-scale code review, test generation, log triage, and batch document work that used to be priced out.
The middle tier is where the competition is now. Neither company competed with its most expensive model this week. Both argued that their mid-priced model reaches the level of a flagship. For engineering teams, the practical default is moving towards “use the mid tier, escalate when needed”, with routing between tiers becoming part of the architecture rather than an afterthought.
Safety is being built into the product. Anthropic says Opus 5.5 was tested before release by external evaluators, including METR, and that it attempts to circumvent boundaries 85% less often than Opus 5. It routes most cybersecurity tasks to an older model unless the user is enrolled in a verification program, and it restricts advanced biology work to approved users. That is a real constraint for security-tooling developers and is worth knowing before migrating. The launch also comes days after CEO Dario Amodei wrote that “fully addressing the risks requires even more prudence”. Some coverage framed shipping a cheaper, faster frontier model the following week as a tension with that statement.
Open questions
- Independent evaluations. Neither lab’s figures have yet been reproduced by a neutral party on a shared harness. Wait for third-party leaderboards before rewriting your routing rules.
- Real-world cost. List prices fell 20% for Opus and 50% for Sol and Luna. The “40% cheaper on typical workloads” and “half the mistakes” claims come from the vendors and depend on the workload.
- Pricing baseline. OpenAI’s developer-forum announcement describes the 50% cut as relative to GPT-5.6 promotional pricing. It is unclear how much of the cut is a permanent price level and how much locks in an earlier promotion.
- The rest of the lineup. Sonnet 5.5 and Haiku 5.5 have not shipped yet. The comparison that matters most for high-volume use, Luna against Haiku, is still to come.
For now, the practical advice is simple: rerun your own evaluations on your own tasks, measure cost per completed task rather than per token, and check how your cache hit rate changes before and after switching.
Sources
- Anthropic — Introducing Claude Opus 5.5
- OpenAI — Introducing GPT-6 Sol and Luna
- OpenAI Developer Community — Announcing GPT-6 Sol and GPT-6 Luna
- TechCrunch — Anthropic releases Opus 5.5 with lower prices and Fable-level performance
- TechCrunch — OpenAI launches GPT-6 Sol and Luna
- Decrypt — OpenAI launches GPT-6 Sol and Luna minutes after Anthropic drops Claude Opus 5.5
- MarkTechPost — OpenAI releases GPT-6 Sol and Luna: 50% cheaper API pricing and benchmarks
- MacRumors — OpenAI’s new GPT-6 Sol and Luna models
- MacRumors — Anthropic launches Claude Opus 5.5
- Yahoo Finance — Anthropic launches Opus 5.5, its first model since Amodei called for AI slowdown