Anthropic has released Claude Opus 5.5, the first model in the Claude 5.5 family that performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. List pricing is $4 per million input tokens and $20 per million output tokens, with cache at $0.20 and speed up 30%. The release matters because it lets businesses standardize on a single model for most tasks instead of splitting work across flagship and economy models.
Fable-level performance at Opus price
List pricing implies a 20% cut from Opus 5 on base rates and a 60% cut on cache, which Anthropic translates into a 40% drop in typical workload cost. Subscription limits have been increased, and a Fast mode is offered at $8/$40 with up to 2.5x speed. Artificial Analysis places Opus 5.5 in the clear lead on intelligence with a score of 58, while its medium, high and xhigh settings sit on the cost-intelligence frontier alongside GPT-6 Luna, GPT-6 Sol and MiMo v2.6 Pro. On max settings the model uses enough tokens to cost slightly more than Opus 5 and only slightly less than Fable 5.1, so cheaper settings often make more sense.
Anthropic positions the model around agentic coding, security and improved communication. Early testers describe longer autonomous runs, higher reliability, cleaner output and steadier back-and-forth in Claude Code. The company says Opus 5.5 puts the most important information up front and follows user writing rules, which helps long sessions stay legible. Visual and spatial work also improved, with BenchCAD Vision2Code at 73% and 96.2% with tools against 95.9% for Astra, Furniture Assembly at 83 against 80, and Chartography at 64.4 against 44.8 for Fable.
The launch contrasts with OpenAI cutting GPT-6 Sol prices by 50% to $2/$10 and Luna to $0.10/$0.50 while asking customers to mix models by task level, whereas Anthropic argues Opus 5.5 fits most tasks. Benchmarks show Opus 5.5 ahead of every model except Astra on many charts and often above Astra, yet Fable 5.1 keeps all seven Vals professional rows plus MedCode, SAGE and both MLCRs. The source reads that cluster as tasks graded purely for correctness, where Fable keeps an edge, while Opus 5.5 leads where presentation and judgment matter. Examples include ProgramBench at 18.5% against 7% for Fable and 5.5% for Astra, Omniscience at 46 against 43, and Terminal-Bench-Science 0.1 at 62% on xhigh against 63% for Astra.
What Opus 5.5 means for business use
For buyers, the practical effect is simpler model routing and lower cost per completed task when medium or high settings suffice. Small teams gain Fable-class coding and document work without maintaining separate pipelines for heavy and light models, while larger organizations can consolidate standards, logging and access controls around one default. Speed gains and higher subscription limits support longer agent runs for refactoring, security review and multi-step analysis. The trade-off is discipline around effort settings, since running everything on max erodes the price advantage.
Limits center on safety controls and uneven strengths. Opus 5.5 ships with zero data retention, which the source questions given stated extremely strong cyber capabilities at least on par with Fable 5.1. Classifiers fire at similar overall rates but more sensibly and with better recovery, though bio and cyber-adjacent work can still hit blocks or unexpected refusals. Testing notes also mention occasional gibberish sentences, an Astra lead on WeirdML v3 at 42.2% against 31.2%, and Astra strength on GDP. pdf. Buyers should verify refusal behavior on their workloads, clarify ZDR terms, and compare quality at specific effort levels rather than assuming max is required.
A useful marker will be whether Artificial Analysis keeps Opus 5.5 on the intelligence-cost frontier through the next two model updates and whether teams report consolidating daily coding on medium and high settings. If Fast mode adoption and higher subscription limits translate into sustained autonomous runs without rising classifier friction, the single-model pitch will hold. If ZDR terms tighten or Fable retains correctness-only tasks, segmentation will persist.
