Anthropic cut Opus 5.5 to $4 and $20 per million tokens, OpenAI answered the same day with Sol at $2 and Luna at $0.10, and the cache read that decides an agent's bill fell to $0.20. On Artificial Analysis's own runs all three reach the cost frontier, and GPT-6 Sol gives back about 100 Elo of professional knowledge work.

What the independent checks show

What the invoice actually charges

Where the cheaper model gives ground

Where your product sits

• Long-running coding and computer-use agents. The cache-read cut is worth more here than the input price is, because these agents replay the same context on every turn. Box tested Opus 5.5 across enterprise knowledge-work tasks and measured 63% fewer tokens used, 42% less verbose output and 30% faster responses than Opus 5 15. Against that, the same model spent the most output tokens per task of anything measured on that index, so measure tokens per finished task rather than per call 12.
• High-volume extraction, summarisation and classification. Luna's $0.10 input and $0.01 cached input put a million-token context in the price band of a cheap open-weight model. Check your own set before moving, because Luna's coding-agent index fell two points and its GDPval Elo fell about 75 8.
• Workflow automation judged per completed task. OpenAI's cost-per-task pitch rests on AutomationBench, where it reports Sol at 33.2% and $0.27 per task against Opus 5's 26.9% at 11.1 times the cost. Those runs compare Sol with Opus 5, which Opus 5.5 replaced hours earlier, so treat the multiple as unproven until someone reruns it 4.
• Anything with a checkable answer and a cheap first tier. A cascade that escalates only the hard cases to a frontier model is the pattern the research literature supports, and it is now cheap to build because the tiers are this far apart 13.
• Overnight evaluation and content generation. OpenAI prices its batch tier at half the standard rate for both new models, and the engineering evidence says moving a workload from streaming to batch costs nothing in output quality 514.

A pilot in two weeks

1. Write the hypothesis with a number in it. "Moving our coding agent to Opus 5.5 at medium effort holds task completion within two points of today and cuts cost per completed task by at least a third."
2. Build the set from your own traffic. 150 real requests with a known-good outcome, including 30 that your current model fails, so any win has to come from somewhere other than the easy cases.
3. Run three arms under one harness. Today's model, Opus 5.5 and GPT-6 Sol, on identical prompts, tools and retry rules. Fix the effort setting in every arm: Opus 5.5 defaults to medium, and comparing it at maximum against a rival at low is the easiest way to misread the result.
4. Count four things. Task completion, cost per completed task, the share of input tokens billed as cached reads, and how often a request escalates to a person. Token price and task cost are different numbers, and the second one is the bill.
5. Set the gate before you read the results. Ship if completion holds within two points, cost per completed task falls by a third, and the escalation rate stays at or below today's.
6. Make the switch a configuration change. Route every model call through a function your team owns, log the model identifier and effort setting on each call, and keep the old route warm for a one-line rollback. Before migrating to Opus 5.5, delete any code that disables thinking or sets a manual thinking budget — the API returns an error for both, and adaptive thinking is now always on 9.

Every figure here comes from an evaluation somebody chose. The index score of 58, the per-task dollar costs, the 119,000 output tokens and the GDPval Elo scores are Artificial Analysis's own runs on benchmarks it selected or built. Anthropic's 66.4% on Terminal-Bench 4.0, its cost-per-task claims and the 40%-cheaper figure are its reporting on its own model, run with production safeguards and with some blocked tasks continued on older Claude models. The Sol and Luna benchmarks and the permanence of their prices come from OpenAI. The 1.9-times-faster and 49%-cheaper result is one consultancy's internal benchmark. METR's evaluation was unpaid but reviewed by Anthropic before publication. The token-verbosity criticism comes from a practitioner on X. The batch and escalation findings come from two arXiv papers, one of which ran inside a single Amazon assistant and the other on call-centre dialogue data. The three charts are Artificial Analysis's own.