The price war of 2026-09-22, and the number inside it that matters: cache reads fell 60%

Source: Simon Willison · The Decoder on GPT-6 Sol/Luna · The Decoder on Opus 5.5 · The Information Raw: raw/rss/2026-09-22-simon-willison-claude-opus-5-5-gpt-6-sol-gpt-6-luna-and-a-new-price-wa.md

TL;DR

On 2026-09-22 Anthropic released Claude Opus 5.5 and, about an hour later, OpenAI released GPT-6 Sol and GPT-6 Luna. Anthropic cut Opus from $5/$25 per million input/output tokens to $4/$20, a 20% reduction, after five consecutive releases held that price. OpenAI halved both new models against their GPT-5.6 equivalents, putting GPT-6 Luna at $0.10/$0.50, one of the cheapest models it has ever shipped. The headline everyone wrote was the price war. The number that matters for anyone building agents is the third column: Opus cache reads fell from $0.50 to $0.20 per million, a 60% cut, three times deeper than the headline cut.

The pricing landscape

GPT-5.6 has a scheduled 25% price increase for November, so GPT-6 is half the price of promotional pricing on the models it replaces. The war is being fought in the tier below the flagships: Astra and Fable 5.1 both sit untouched at $10/$50.

Why the cache-read line is the real story

Simon Willison states the mechanism plainly: in longer agentic conversations, 90%+ of input tokens are processed at cached token prices. So for an agent workload, the cached-input column is most of the input bill, and Opus 5.5 cut it by 60% while cutting the headline input price by 20%.

That is a deliberate pricing signal about where the load is. Anthropic is not discounting inference, it is discounting prefix reuse, which is the specific thing an agent loop does hundreds of times per task. The wiki has been treating the KV cache as a technical object with compression ratios and eviction policies. This is the 08-14 kv-cache entry, "the cache becomes a billing surface," maturing into the phase where the billing surface is being used as a competitive weapon and is moving faster than the compute price around it.

How this connects to the rest of the wiki

It validates the harness-cost thesis with a vendor's own price sheet. HarnessTax (09-22) found that across 21 model-harness combinations, cost per attempt spanned $0.67 to $1.33 while success spanned 96.7% to 97.8%, and traced the spread to standing context: longer instructions and larger tool definitions ride along on every call, so a heavy harness pays its fixed overhead once per turn. Standing context is exactly what a cached prefix is. A 60% cut in cache-read price cuts the harness tax by 60% for every heavy harness, which narrows the gap HarnessTax measured without anybody changing a line of harness code.

It cuts against the Meta-Harness and RRSI optimization case, slightly. Today's Stanford/MIT Meta-Harness reports discovered harnesses beating hand-engineered agents on TerminalBench-2 while cutting context token usage 4x, and RRSI (09-22) produced a harness running on 30% fewer policy tokens. Those are token-count optimizations whose dollar value just fell by 60% on the cached portion. The optimizations still matter for latency and for context-window pressure, but the economic argument for them weakened on the same day the papers landed.

The 09-21 prediction that "price was never the binding constraint, calibration is" survives contact with this. The llm-routing page recorded that claim two days ago, on evidence that the crudest possible routing policy (send everything cheap) already captures most of the available saving. A 50% price cut on the cheap tier and a 60% cut on cache reads makes the cheap-by-default policy cheaper still, which strengthens rather than weakens the finding: if the market keeps cutting prices faster than routers can learn, a learned router's payback period keeps receding.

The economics are moving in one direction and the infrastructure costs in the other, which is the tension to watch. The same week: data-centre bonds on a Jane Street-leased build yielding 11.3% against an August issue two points tighter; CoreWeave completing an upsized $4.2 billion convertible; Anthropic in talks to lease up to 1 gigawatt directly from Apollo-controlled Stream Data Centers to reduce cloud dependence, filling it with Broadcom-and-Google-designed TPUs; the CFTC stalling CME's GPU rental futures; roughly $130 billion of US data-centre projects blocked or delayed in Q1 2026 with 71% of surveyed Americans opposing one near them; and Texas freezing new grid-connected projects. Token prices are falling while the cost of the capacity to serve them is rising and its financing is getting more expensive. Those cannot both continue indefinitely.

The quality caveat nobody should skip

The Decoder's read is that GPT-6 Sol and Luna "barely move the needle on performance," and independent analyses found little gain in actual intelligence: this is a price cut, not a capability release. And Willison recorded a concrete regression on Opus 5.5 at "max" thinking level, which failed to return a response at all on his standard SVG test, twice, because it exhausted the 128,000-token output limit while still reasoning. Two failures at $2.56 each and nearly 20 minutes. His conclusion, that max is "effectively useless" if it over-thinks to breaking point on a trivial prompt, is a data point directly supporting Taste-Bench's finding today that a larger reasoning budget does not improve decision quality. Two independent observations on the same day that the effort knob has a ceiling and possibly a cliff.

Related pages

• compute-economics · kv-cache · llm-routing
• HarnessTax · Taste-Bench