Anthropic launches Claude Opus 5.5 as OpenAI releases GPT-6 Sol and Luna, intensifying API price war

Anthropic released Claude Opus 5.5, calling it the first model in the new Claude 5.5 family. About an hour later, OpenAI released GPT-6 Sol and Luna. On the same day, both companies put "price cuts" at the center of their launches.

Anthropic's official line is that Opus 5.5 reaches Claude Fable 5.1 levels on most tasks and runs at 40% lower cost than Opus 5. On OpenAI's side, launch prices are 50% lower than GPT-5.6. These two numbers made the day's most eye-catching headlines, but they come from the vendors' own statements—they are claims, not independent verification.

The hard number that can be confirmed is token list pricing: Opus 5.5 dropped from $5/$25 per million input/output tokens to $4/$20, a 20% reduction. This is a change at the price-list level, which is not the same thing as "40% cheaper per task."

The claim of "about 30% faster and about 40% cheaper per task" comes from ClaudeDevs and lydiahallie, representing the vendor and its employees. It describes task-level cost, and task-level cost is affected simultaneously by list price, token usage, and cache pricing—which is exactly where the divergence appears later.

Artificial Analysis's measurement is the most critical counter-evidence in this launch. At the max tier, Opus 5.5 costs $5.98 per Intelligence Index task, close to Opus 5 (max) at $5.86. In other words, at the highest compute tier, the per-task savings relative to Opus 5 disappear.

Their breakdown lays out the reason: rising token usage alone increases per-task cost by about 80%, to $10.51; the 20% base price cut brings it down to $8.41; cheaper cache reads ($0.20) bring it down further to $5.98. The price cut is real, and the usage increase is real—the two nearly cancel each other out.

From this one can infer (this is an inference, not a vendor statement): the so-called "40% cheaper" applies to the default medium setting, not the max tier. The two framings cannot be directly compared—one describes task cost under the default configuration, the other measures the highest compute tier.

Opus 5.5 has become the default model for Claude Code and the Claude app (including Cowork), with a default effort of medium. This means what most users actually encounter is precisely the tier that corresponds to "40% cheaper."

Changes on the subscription side are equally concrete: the 5-hour session limit increased by 20%, the price cut gives 25% more usable capacity within the limit, and Pro, Max, and Team users receive limit reset credits they can use at a time of their own choosing. On the availability side, Opus 5.5 has launched in Claude Code, the Claude Platform API, and Claude Tag for Slack.

Third-party and partner evaluations present a picture with both wins and losses. On Vals Index, Opus 5.5 ranks first, up 2 places and 2 points from Opus 5. On FrontierSWE (Proximal) it scores 62.3%, placing second, behind GPT-6 Astra's 65.5%, but above Fable 5.1's 56.3% and Opus 5's 52.0%.

On FrontierCode 1.1 (Cognition) Extended it scores 65.3%, taking first place; on CursorBench it scores 57.8% (Max), becoming the new top result; on ParseBench table tasks it scores 93.9%, a 7-point improvement over Opus 5.

These results need to be read with their configurations in mind. Vals's RSI evaluation was run under specific settings: native Claude Code, max effort, 1M context, 128K max output, temperature 1, and the results depend on that configuration. ParseBench's negative conclusions—charts and layout remain weak, OCR costs too high—come from LlamaIndex, a vendor with a competing product, and its position needs to be factored in.

Anthropic says Opus 5.5 is the first Opus with Fable 5.1-level cyber, bio, and frontier LLM development safety safeguards, that flagged requests fall back to another model, and that it is reducing false positives. The actual frequency of fallbacks and the false-positive rate remain unclear.

On the roadmap, Anthropic says Sonnet 5.5 and Haiku 5.5 will follow "in the coming weeks," which refutes rumors that Haiku was being discontinued; but the specific release timing and specifications remain unclear.

Writing improvements are another thread highlighted separately: Latent Space says Opus 5.5's writing improvements are notable and has already migrated AINews to the model immediately. This is the source's own user experience, not a benchmark score.