Claude Cost, Context, and Latency: The September 2026 Model Tradeoff Report

The rate card is only the first line

Anthropic's current model table places Fable 5.1 at $10 input and $50 output per million tokens, Opus 5 at $5 and $25, Sonnet 5 at a standing $2 and $10, and Haiku 4.5 at $1 and $5. GPT-5.6 preview pricing is listed by OpenAI as Sol $5 and $30, Terra $2.50 and $15, and Luna $1 and $6 per million tokens.

Two of those numbers changed since the July baseline of this site. Opus 5 replaced Opus 4.8 at the same $5/$25 rate, and Sonnet 5 stayed at $2/$10 instead of rising to the previously announced $3/$15. The second is covered in detail in the Sonnet 5 pricing report.

Those numbers are useful, but agent workloads rarely map cleanly to "one prompt in, one answer out." Long-running coding agents read repositories, update context, call tools, retry, produce patches, and sometimes summarize their own traces. Output-heavy models can become expensive even when input price is moderate.

Cache reads are now a tier-level differentiator

Claude cache reads are priced at 10% of base input on most models, but Fable 5.1 and Mythos 5.1 are priced at 2.5%, or $0.25 per million tokens. Anthropic's launch announcement estimates roughly 25% savings on typical workloads and up to about 45% on highly agentic ones as a result. The Batch API remains 50% off for delay-tolerant work.

This breaks the intuition that the price ladder is a simple multiple. On uncached input, Fable 5.1 costs twice Opus 5. On cached input, Fable 5.1 reads at $0.25 per million against Opus 5 at $0.50 — the ordering inverts. For an agent whose per-turn prompt is mostly a stable repository summary or system prefix, that inversion can be the dominant cost term.

The practical consequence is that a cost comparison built only from rate-card input and output prices will misrank the Claude lineup for exactly the workload shape that most production agents have. Model the cache hit rate explicitly or the estimate is not usable.

Context window is capacity, not free value

Anthropic lists Fable 5.1, Opus 5, and Sonnet 5 with 1M-token context windows and 128K max output. Haiku 4.5 is listed at 200K context and 64K output. Large windows make the top three attractive for large repositories, document bundles, and multi-turn agents. They do not mean every task should carry a million tokens.

A better cost plan separates static context from dynamic task state. Prompt caching helps when many runs reuse the same instructions or repository summary — which, given the cache-read rates above, now also influences which model you pick. Shorter context can still win when the relevant files are retrieved precisely.

Latency, effort, and the procurement test

For current Claude models, effort trades intelligence against latency and cost, and adaptive thinking is always on for Fable 5.1. Defaults vary by surface: the platform docs list high default effort in Claude Code and medium on Claude.ai and Cowork. OpenAI's GPT-5.6 Sol preview similarly adds max and ultra modes. Model comparison in 2026 is less like choosing an engine and more like choosing an engine plus its operating point.

For each candidate model, measure total cost per accepted result. Include failed attempts, human-review time, extra tool calls, and retries after refusal or policy routing. A model with a higher token price can be cheaper if it finishes in fewer turns and needs less human repair.

Anthropic's own guidance supports a cheap-first ladder: start with Opus 5, escalate to Fable 5.1 only when evaluations at higher effort still fall short. A rational test compares Sonnet 5 at a named effort against Opus 5 at a named effort on the same budget cap, with caching configured the way production will configure it.

FAQ

Cite this page