Price used to be the first filter when choosing a model. As of this autumn, that filter no longer works.

Between September and October, Anthropic, OpenAI and Google all shipped new generations within five weeks. After the refresh, pricing at each tier lines up almost exactly.

Three columns, identical numbers

Frontier tier: Claude Fable 5.1 and GPT-6 Astra both list $10 per million input tokens and $50 per million output.

Workhorse tier: Claude Sonnet 5.5, GPT-6 Sol and Gemini 4 Argon all sit at $2 and $10.

Entry tier: Claude Haiku 5.5 and GPT-6 Luna both at $0.10 and $0.50.

Three months earlier the same table showed real spread: Sonnet 5 at $3/$15 after promos, GPT-5.6 Terra at $2.50/$15, Gemini 3.1 Pro at $2/$12. The cheapest tier ran from $0.25/$1.50 up to $1/$5. Now the small models have converged at a tenth of the old floor — a tenfold cut for Anthropic's smallest model — and the workhorse tier lands on the same $2/$10 everywhere.

Why convergence happened

Three forces stacked up.

The competitive structure stabilized. With capabilities close, any unilateral price increase loses deals immediately.

Cost structures converged. Same accelerators, same inference optimizations, compressed unit-cost differences.

Pricing strategy shifted to developer acquisition. The API is the funnel; the model does not have to be profitable on its own.

Unit price stopped being informative

When all three charge the same, price stops telling you anything. What decides the bill:

Tokens per task. A model that reasons less and gets it right the first time can be cheaper at a higher rate. Anthropic's pitch for Opus 5.5 is exactly this: close to flagship quality at roughly 40 percent lower running cost, partly from pricing and partly from fewer tokens per step.

Retry rate. A ten-point gap in first-pass success can double effective cost.

Cache hit rate. Long system prompts and fixed knowledge bases are where caching either works or costs you.

Batch discounts. Non-realtime work batched typically halves the bill.

How this reads against Chinese models

Domestic models price more aggressively, some below the international entry tier. Comparing list prices alone leads to the wrong conclusion, though — differences in long-context handling and tool-calling stability show up directly as retries.

A selection checklist

1. Run A/B tests on your real tasks, recording tokens consumed and success rate per task.
2. Compute cost per completed task, not price per million tokens.
3. Factor in caching, batching and context compression as engineering levers.
4. Keep two vendors and route by task type rather than switching wholesale.

Price convergence is a signal: LLM APIs are becoming a utility. That is good for buyers, and it moves competition to engineering — who extracts more useful output from the same dollar.