A million output tokens from a frontier language model costs somewhere between $0.50 and $50 this month, depending on which vendor you ask. That is a 100-fold gap on published rate cards, and the number that lands in your cloud bill depends less on model quality than on where you sit on the vendor ladder.
ActuIA, a European AI news outlet, surveyed the public pricing pages of OpenAI, Anthropic, Google, Mistral, DeepSeek and xAI on October 6, 2026, along with Amazon Bedrock and Google Cloud. The sample covers 18 offers. Output prices span from 50 cents for OpenAI's GPT-6 Luna to $50 for both GPT-6 Astra and Anthropic's Claude Fable 5.1. See our AI coverage for background on this pricing race.
The cheap end of the table has gotten genuinely cheap. GPT-6 Luna prices input at $0.10 per million tokens and output at $0.50. Mistral Small 4 sits at $0.15 input and $0.60 output. DeepSeek's off-peak deepseek-flash runs $0.15 input and $0.60 output, falling to $0.03 per cached million at off-peak hours. These are API prices, not marketing tiers buried in a free trial.
The expensive tokens are expensive for a reason vendors state but developers dispute. Anthropic charges $10 input, $50 output for Claude Fable 5.1 and argues cache reads make up the bulk of agentic and coding costs. OpenAI's GPT-6 Astra matches that $10/$50. Anthropic's newer mid-tier models, Claude Opus 5.5 and Claude Sonnet 5.5, cut both by half or more, down to $4/$20 and $2/$10 respectively.
Why September Was the Busiest Month for Rate Cards
The October snapshot did not arrive in a vacuum. Every major vendor posted a new or repriced model between September 1 and September 29. According to ActuIA's month-by-month log, Anthropic launched Claude Fable 5.1 on September 1 and lowered cache reads to $0.25, OpenAI released GPT-6 Astra on September 3, Google shipped Gemini 3.8 Flash on September 2, and DeepSeek dropped V4.1 Flash prices on September 10. The month closed with Claude Sonnet 5.5 on September 28 and a halved cache read price for GPT-6.1 Sol on September 29.
Three of those September launches were billed as cache read reductions, not output price cuts. Fable 5.1 and GPT-6.1 Sol changed only that line. Opus 5.5 cut cache reads by 60 percent while pulling input and output down 20 percent. Anthropic has said publicly that cache reads are the token line driving most agentic and coding spend. The vendors are competing on re-read cost rather than on first-pass output cost.
One backward-looking deadline still matters. Anthropic had planned to raise Claude Sonnet 5 from $2/$10 to $3/$15 effective September 1. Instead, on August 10 it made the introductory $2/$10 rate the standard price. The increase never took place. That single reversal remains the most direct price cut of the summer for Sonnet customers.
Cache Reads Now Carry the Margin
Prompt caching exists because an agent conversation re-sends the same context every turn. Once the first turn pays full input price, every subsequent turn reads from cache at a reduced rate. ActuIA's table shows cached input running 1 percent of full input price at OpenAI, 2.5 to 20 percent at Anthropic depending on tier, and 10 percent at Google.
That math now dominates developer bills. A coding agent with a 100,000-token codebase in context reads most of its input tokens from cache on every turn. Cutting the cache rate by half moves a team's API bill more than cutting the list output rate by the same percentage, because the cached volume is far larger.
Long-context requests quietly inflate prices too. OpenAI charges higher rates above 272,000 input tokens. Google indexes price on prompt size, climbing from $2/$12 to $4/$18 for Gemini 3.1 Pro above 200,000 tokens. xAI bills the higher rate on all tokens in the request once the 200,000-token threshold is hit rather than on the marginal tokens. A developer who does not track prompt length can blow past a rate card without noticing.
Deadlines That Already Sit on the Rate Cards
Three pricing dates are already published and already affect planning. Google's Gemini 3.8 Flash introductory price of $0.75/$3.75 expires December 31, 2026, doubling to $1.50/$7.50 on January 1. The GPT-5.6 Sol promotion holds until at least November 21, 2026. Anthropic will retire Claude Sonnet 4.5 from its API on November 30.
Anthropic also changed policy without touching prices. On October 8 it updated its usage policy to prohibit sustained and needless abusive behavior toward Claude, effective November 12. The change is not a price change and does not alter API rates, but it does matter for teams running automated red-team evaluations, because the policy does not publish a detection threshold. There is no new tariff attached to rude prompts.
What the Spread Means for Buyers
The 1-to-100 range is not a sign that one vendor made a mistake. It is the market segmenting itself. At the bottom, Luna, Gemini Flash-Lite and Mistral Small 4 sell at pennies per million tokens for high-volume low-stakes work. In the middle, Sonnet 5.5, GPT-6.1 Sol and Gemini 3.1 Pro compete at $2/$10 for serious agent work. At the top, Fable 5.1 and GPT-6 Astra charge $10/$50 for the most capable models.
The open question is whether the top tier can hold. ActuIA's own earlier analysis noted that new mid-tier models match last quarter's flagship performance at half the price. If the gap between the middle and top tiers is only a few months of capability, the $50 token is a subscription rather than a market price.
Buyers should not read the October table as stable. Vendors published six releases in three weeks and are still working through promotional and introductory rates. The vendor's rate card remains authoritative, and ActuIA is careful to note that its own table is a snapshot taken between 00:00 and 01:00 UTC on October 6. Anyone budgeting from it should index on the vendor page and recheck before committing capacity.
The next rate card survey runs October 13. Between now and then the GPT-5.6 Sol promotion approaches its November expiry and Gemini 3.8 Flash's January price doubling gets one month closer. Developers running cost models on Anthropic alone should track the Sonnet 4.5 retirement on November 30, since any workload still on that model needs a migration plan before the deadline.
ActuIA's full survey, including the euro-conversion table using the European Central Bank's October 5 reference rate of 1.1204, is at ActuIA. The vendor rate cards are the authoritative source; ActuIA publishes the cross-vendor comparison because no single vendor will print a side-by-side table of its competitors.