Anthropic says Opus 5.5 beats OpenAI's top model, GPT-6 Astra, on coding. OpenAI says Sol beats Claude Opus 5 on business workflows for under a tenth of the cost per task.
Anthropic released Claude Opus 5.5, the successor to Opus 5. OpenAI released GPT-6 Sol and GPT-6 Luna, which replace its GPT-5.6 models of the same names and rank below Astra. The two announcements came about 90 minutes apart.
Opus 5.5 costs $4 per million input tokens and $20 per million output. Opus 5 cost $5 and $25. Cache reads, which Anthropic says make up most of the cost of agentic and coding work, fall 60% to $0.20. Because the model also spends fewer tokens per task, Anthropic estimates that typical workloads cost about 40% less than on Opus 5, with output more than 30% faster.
OpenAI says Sol and Luna cost 50% less than the promotional API rates of the GPT-5.6 models. Sol now costs $2 and $10 per million input and output tokens, down from $4 and $20. Luna drops to $0.10 and $0.50, from $0.20 and $1.20. The company credits the cut to better caching and inference, and bills cached input tokens at a 90% discount.
Anthropic's own launch table gives Opus 5.5 66.4% on Terminal-Bench 4.0, a coding test run at the command line. Astra scores 57.9% on the same test, and Fable 5.1, the pricier tier above Opus in Anthropic's lineup, scores 55.8%. Astra leads on terminal-based scientific research, 64.6% to 58.7%. On most work, Anthropic puts Opus 5.5 level with Fable 5.1 at two-fifths of Fable's input and output rates.
On AutomationBench, a test of business workflows across apps, Zapier's leaderboard has Astra first at 41.4% and Opus 5.5 second at 40.0%, both at maximum reasoning effort. One setting down, Opus 5.5 scores 34.4% for $0.80 per task and Sol 33.2% for $0.27. OpenAI compares Sol's 33.2% with 26.9% for the older Opus 5, which it says costs more than eleven times as much per task.
Anthropic says Opus 5.5 is strong enough in cybersecurity and biology to be misused. Most security prompts therefore go to the older Opus 4.8 and flagged biology prompts to Opus 5, the same safeguards it put on Fable 5.1.
Opus 5.5 runs in Claude's apps, on Anthropic's API and on the major cloud platforms. Anthropic also raised the five-hour usage limits on most Claude plans and gave subscribers one reset of those limits to use whenever they choose. Its Sonnet 5.5 and Haiku 5.5 follow in the coming weeks.
Sol and Luna are in ChatGPT Work, Codex and OpenAI's API, and Luna is also open to free users in the desktop app.
In an independent index from Artificial Analysis, which combines ten evaluations, Opus 5.5 scores 58 and Sol 48. A task in that index costs an average of $5.98 to run on Opus 5.5 and $1.06 on Sol.
