[AINews] Claude Opus 5.5, the new default model for AINews — and everybody cuts prices 40-50%

What it claims

Anthropic positions Claude Opus 5.5 as Fable 5.1-level capability at Opus pricing. It is said to be about 30% faster and about 40% cheaper per task than Opus 5. List price drops 20%, from $5/$25 to $4/$20 per 1M input/output tokens. Cache reads now cost $0.20. The "40% cheaper" figure holds at the default medium effort; at max effort the per-task saving largely disappears. OpenAI's same-day GPT-6 Sol costs $2/$10 and Luna costs $0.10/$0.50, each about 50% below its GPT-5.6 predecessor. That makes Sol half Opus 5.5's list price. AINews switched its own pipeline to Opus 5.5 after a side-by-side run against Sol, citing writing quality.

Evidence

Cost per task at max effort (Artificial Analysis Intelligence Index): Opus 5.5 costs $5.98 against $5.86 for Opus 5 (max). Their decomposition:

• Higher token usage alone would raise cost by about 80%, to $10.51.
• The 20% base-price cut brings that to $8.41.
• Cheaper cache reads bring it to $5.98.

Vals reports that the extra token use is heaviest on coding, which is also where the largest gains are.

Third-party results:

Vals ran RSI in native Claude Code at max effort, with 1M context, 128K max output and temperature 1.

Effort-scaling anomaly. On one agentic-coding chart, xhigh effort costs about 2.8x more than medium and scores 3.2 points lower. The explanation offered: FrontierCode penalizes unnecessary changes, and higher effort produces scope creep. Opus 5 showed the same pattern.

OpenAI's comparisons are framed as cost per task. They claim Sol at xhigh beats Opus 5 max on AutomationBench at about 9% of the cost. Devin reports Sol matching GPT-5.6 Sol at 61% lower cost per task.

How to use it

• Where it runs: Claude Code, the Claude Platform API, the Claude app (including Cowork) and Claude Tag for Slack. It is now the default in Claude Code and the app, at default effort medium.
• Prompt cache: switching effort mid-session does not break the cache on Claude Code v2.1.280+.
• Anthropic's prompting changes:
• Hand over the whole task, and define what "done" means and where to check in.
• Drop "think carefully", because the model always thinks first.
• After a long run, ask what it needs to go further.
• Old prompt tricks may conflict with the new training.
• Subscription limits: 5-hour session limits are up 20%. Pro, Max and Team users get a banked rate-limit reset.
• Coming next: Sonnet 5.5 and Haiku 5.5 are due "in the coming weeks".
• OpenAI side: caching gives up to 90% discounts on cached input reads. There is also a new Prompt Caching Dashboard and a diagnostics API for broken cache reuse.

Limits

• Safeguard fallback. This is the first Opus with Fable 5.1-class safeguards on cyber, bio and frontier LLM development. Flagged requests are silently routed to another model, and Anthropic acknowledges incorrect flags. One user reported a fallback on a benign request. A single test suggests the frontier-LLM-development classifier targets Chinese hardware; this is unverified.
• ProgramBench. The system card's near-100% figure comes from a 166/200 subset that likely excludes the hardest programs, such as FFmpeg and the PHP compiler. It also reports average test pass rate rather than full task completion. Partial solves often pass 60–70% of tests, which inflates the pass-rate number.
• Known bug: the live model sometimes generates user turns.
• ParseBench: the model is still weak on charts, formatting and layout. LlamaIndex calls 5.8¢/page too expensive for production OCR; note that LlamaIndex sells a competing product.
• AI R&D: one reading of the system card says the model is "roughly similar" to its predecessors here, with the gains concentrated in agentic coding.
• Unconfirmed claims: model size (said to be smaller than Opus 5) is speculation. Arena has no scores yet.
• Before relying on the numbers: most figures are vendor or partner claims relayed through tweets, and the newsletter states it compiled them rather than reproducing them.