The token price war hits $0.25 per million: what cache-read cuts mean for agents
In the first four days of September 2026, Anthropic, Google, and OpenAI all moved on the same line item: cached-context pricing. For long-lived coding and research agents, the cost of re-reading context — not the headline token price — is now the deciding economic variable.
Three labs moved on API pricing in the space of about 72 hours — and none of them touched the headline numbers. The fight has shifted to a single line item on the invoice: the price of re-reading cached context. For agentic workloads, that is the line item that decides the bill.
What happened, day by day#
- September 1 — Anthropic. Claude Fable 5.1 launched with its base price unchanged at $10 per million input tokens and $50 per million output tokens. The only price that moved was cache reads: cut 75%, from $1.00 to $0.25 per million tokens. Anthropic estimates the cut lowers typical bills by about 25% and highly agentic workloads by up to roughly 45%, based on four weeks of its own August usage data.
- September 2 — Google. Gemini 3.8 Flash went generally available, holding the promotional rate of $0.75 input / $3.75 output per million tokens through December 31, 2026 — with cache reads at $0.075 per million. Google's third Flash release in six weeks is explicitly positioned for software engineering and autonomous agents. The catch: on January 1, 2027, those rates double to $1.50 / $7.50.
- September 3–4 — OpenAI. GPT-6 Astra launched at the exact same sticker as Fable 5.1 ($10 input / $50 output per million), with cached input reads at $1 per million and automatic prompt caching built in. No separate cache-read discount below the standard 90% off — but the long-context structure is novel: above roughly 272,000 input tokens, the entire request reprices, with input doubling to $20 and output rising to $75.
Same week, three vendors, three different bets on the same question: how should the repeated, re-read context of an agent session be priced?
Why cache reads are the whole game for agents#
Prompt caching lets a provider reuse the KV cache of a stable prompt prefix instead of recomputing it. Re-sends of the same system prompt, tool definitions, repository state, or accumulated transcript are billed at the discounted cache-read rate. A coding agent that loads a large context once and works against it for hours will re-read that prefix on every turn — so the cache-read price, not the input price, becomes its effective marginal cost.
Standard industry convention prices cache reads at about 10% of the input rate. Anthropic broke that convention: Fable 5.1 reads cache at 2.5% of its input price ($0.25 against $10). That produces an odd inversion inside Anthropic's own lineup — Fable 5.1's standard input and output rates are twice those of Claude Opus 5 ($10/$50 versus $5/$25), yet its cache reads are cheaper than Opus 5's ($0.25 versus $0.50 per million). The structure deliberately favors long-lived sessions over stable context, which is exactly the shape of a coding agent mid-assignment.
A concrete illustration from Anthropic's launch coverage makes the arithmetic tangible: a workload with 100 million fresh input tokens and 900 million cached tokens costs $1,900 in input charges at the old $1 cache-read rate, and $1,225 at the new $0.25 rate — a 35.5% reduction in input cost. Output charges are unchanged, so the savings percentage shrinks on a full invoice, but for read-heavy agents the input line dominates.
Three structures, three assumptions#
The comparison between Fable 5.1 and GPT-6 Astra is the sharpest lens on the price war, because the headline prices are identical. Side by side:
| Claude Fable 5.1 | GPT-6 Astra | |
|---|---|---|
| Input / output per 1M tokens | $10 / $50 | $10 / $50 |
| Cache reads per 1M tokens | $0.25 | $1.00 |
| Long-context treatment | No surcharge published | Whole request reprices above ~272K input tokens |
| Caching mechanics | Explicit (cache breakpoints) | Automatic |
Anthropic is betting that agent workloads concentrate in long sessions over stable context, and has priced the dominant cost of that pattern down aggressively. OpenAI is betting that a very large context window is a premium capability: the million-token window exists, but the pricing says use the first quarter of it routinely and the rest deliberately. Neither is wrong — they are different predictions about where workloads land.
Google's bet is different again: raw undercutting on the headline rate rather than on the cache-read line. At $0.075 cache reads, Gemini 3.8 Flash is the cheapest repeated-read option of the three — but the promotional clock is the catch. Teams budgeting for 2027 at 2026's Flash prices will face a 2x repricing, and Google's own guidance notes the model "works harder" on complex tasks, producing more tokens that partially offset the low per-token price.
What to watch before you switch#
- Cache-hit rate beats model choice. If most of your billed input is re-read prefix, the cache-read price is your real price and the headline rate is close to irrelevant. Measure your own hit rates before migrating; poorly structured prompts — reordered tools, edited early context — silently destroy cache reuse.
- Output can erase the discount. Early third-party measurements of Fable 5.1 diverge for a reason: the model reportedly emits roughly 1.7x more output tokens at high effort levels, and output bills at $50 per million. Anthropic's own customer launch story has Cognition reporting roughly a 54% per-task cost drop on coding benchmarks — but that is one workload shape, not a universal outcome. One write-heavy agent and one read-heavy agent can show opposite results on the same model.
- Watch the cliffs, not the averages. GPT-6 Astra's long-context threshold reprices the whole request, not just the overage. What matters is how often your request distribution crosses 272K input tokens, not your mean prompt length.
- Price real traffic. Take a representative day of requests and cost it under each structure. This takes an afternoon and routinely changes the answer — as two teams on the same model can carry wildly different effective rates depending entirely on workload design.
- Recheck when your product changes. Adding memory, retrieval, or tool use changes your token shape, and the pricing structure that suited you at launch may not suit you six months later.
The takeaway#
The September price moves are less a price war than a divergence in pricing philosophy. Anthropic is subsidizing the specific act of an agent remembering — re-reading its own context across long sessions — because that is the unit of cost that agentic workloads actually generate. OpenAI is charging for memory at a premium past a threshold. Google is discounting everything for four months to buy adoption. All three agree on the same underlying fact: the frontier model business is now an agent-infrastructure business, and the cache-read line is where agent economics are won or lost. Build your cost models around it, not the sticker price.