Anthropic released Claude Opus 5.5 and cut its price 20%, and minutes later OpenAI put out two new GPT-6 models, Sol and Luna, at half what their predecessors cost. Input on Opus 5.5 runs $4 per million tokens, output $20; Sol comes in at $2 and $10, Luna at 10 cents and 50 cents. Within a single day the two largest model makers moved the price of frontier-class work down by a fifth to a half — this matters for any company that budgets for AI by the token.
What the two launches actually include
Anthropic said Opus 5.5 matches its Fable 5.1 model on most work, with output generation more than 30% faster than Opus 5, which shipped in late July. Cache reads dropped 60% to 20 cents per million tokens, and a faster serving mode carries a premium at $8 input and $40 output. On Terminal-Bench 4.0, an agentic coding test, the new model scored 66.4% against 55.8% for Fable 5.1; on AutomationBench it reported 40% task completion where Opus 5 managed 26.9%. The widest gap sits in scientific research: Terminal-Bench-Science 0.1 gave Opus 5 a score of 29% and Opus 5.5 a score of 58.7%. On GDPval-AA v2.1, a knowledge work test scored in Elo, the new model rated 1846 against 1708. One tester finished a 680,000-line code migration in under a day, work Anthropic said would have taken an engineering team weeks.
OpenAI aimed its two releases lower down the cost curve. Sol and Luna were built on much the same training methods as GPT-6 Astra, which the company began rolling out Sept. 3, then tuned for cost; Astra for Law, a configuration built for legal research, arrived Sept. 17. Luna is pointed at high-volume routine jobs such as summarization and extraction, Sol at recurring coding and agent work. OpenAI reported Sol completing 33.2% of tasks on AutomationBench at 27 cents apiece and 68.8% on the DeepSWE v1.1 software engineering test, within 1.1 points of the earlier Claude Fable 5; Luna reached 66.6% on the same test. Sol makes about half as many mistakes as its predecessor, and Luna at higher effort settings matches GPT-5.6 Sol at roughly a hundredth of the cost. OpenAI is discounting reads of cached input tokens by 90% and said coding agents can now change reasoning effort and their tool set without breaking the cache.
Both companies published customer testing notes with the release. Yashodha Bhavnani, vice president of AI products at Box Inc., said answers from Opus 5.5 were 40% less verbose without losing accuracy; GitHub Inc. Chief Product Officer Mario Rodriguez said that in VS Code the model solved more terminal tasks than Opus 5 in less than half the steps. Anthropic called Opus 5.5 the strongest performer to date on the automated behavioral audit it uses for alignment testing, and said attempts to circumvent containment boundaries dropped 85% against Opus 5. Restrictions on cybersecurity work carry over from Fable 5.1, with most such tasks routed to the older Opus 4.8, and high-risk biology work fenced off behind verification programs that now include a life sciences track. Frontier Design and METR both tested the model before release.
What this means for companies buying AI
The practical effect is a wider usable price band. A routine summarization or extraction pipeline can now run on Luna at 10 cents and 50 cents per million tokens, while coding and agent workloads can stay on Sol at $2 and $10 — both cheaper than the GPT-5.6 versions they replace. Anthropic's 60% cut on cache reads changes the economics of long-running agents that repeatedly re-read context: repeated calls over the same prompt cost a fifth of what they did before. For a small company this shows up as a lower monthly bill at the same volume; for a large one it means workloads previously held back by cost, such as bulk code migration or continuous test runs, can be moved into production without a fresh budget cycle.
Several conditions still need checking before a switch. OpenAI's comparisons run against Fable 5.1 and Opus 5, so neither launch offers a direct head-to-head between Opus 5.5 and Sol or Luna. Anthropic's faster serving mode doubles the price to $8 and $40, which erases part of the discount if latency targets require it. Cybersecurity and high-risk biology tasks are not available on Opus 5.5 on the same terms as general work, and most cybersecurity requests are routed to Opus 4.8. On the OpenAI side, Sol and Luna are live for developers as gpt-6-sol and gpt-6-luna and in ChatGPT Work and Codex for paid tiers, but neither has reached the main ChatGPT chat surface yet, so consumer-facing products cannot rely on them. Claude Sonnet 5.5 and Claude Haiku 5.5 are due in the coming weeks and may shift the price ladder again.
The marker to watch is the rollout of Claude Sonnet 5.5 and Claude Haiku 5.5 in the coming weeks, together with whether Opus 5.5 pricing holds once those tiers arrive. If the mid-tier models land at or below Sol's $2 and $10 while keeping the benchmark gains, the discount becomes the standard rather than a launch promotion, and AI cost assumptions in 2027 budgets will need rewriting.
