The API pricing log

A dated record of changes in published API prices for frontier language models. A vendor’s pricing page shows today’s rate and nothing else; this page keeps the history. What moved, when, and by how much.

Last updated September 21, 2026. New entries are added as changes are verified.

Entries

How entries are made

Primary sources only. Every entry comes from a vendor’s own pricing page or announcement. An entry is dated by when the change took effect where the vendor names that date, and marked “read” with our verification date otherwise.

Aggregators are not list prices. A reseller’s “cheapest price” reflects routing across providers and can move without the vendor changing anything. Useful signal, but a different quantity; it is never recorded here as a vendor price change.

The vendor names the predecessor where it can. A new generation is logged as a model evolution, priced against the model it replaces. Which model that is comes from the vendor: where its own documentation maps a new model to an earlier tier, that pairing is the one used here, even where our reading of the line-up would have chosen differently. Where a vendor says nothing, the pairing tracks the role a model plays, the frontier model or the cheap workhorse, rather than its label, because names and ladders change constantly.

Corrections are marked, not silently edited. Where an entry has changed since it first published, the change and its date appear on the card.

Per-token price is not cost per task. Token efficiency and tokenizer changes move real cost independently of the headline rate. Changes of that kind are logged as effective price changes.

Rates are re-verified, not trusted. The DeepSeek entry surfaced in a routine re-check of a rate that had been verified only three weeks earlier; nothing on the vendor’s page announced the change.

Prices are per million tokens; a change reads as before → after. Newest first.

Effective

DeepSeek’s cheap slot moves to V4.1-Flash at a lower price, and V4-Pro is folded into it

Model evolutionCheap slot
New modelPredecessorNewChange
deepseek-flash (V4.1)succeeds deepseek-v4-flashpeakInput$0.44→$0.30-32%
Cache hit$0.014→$0.006-57%
Output$1.32→$1.20-9%
off-peakInput$0.22→$0.15-32%
Cache hit$0.007→$0.003-57%
Output$0.66→$0.60-9%

DeepSeek released V4.1-Flash on September 10 under the name deepseek-flash and retired V4-Flash; requests to the old name are served by the new model at the new rate. At peak, input falls 32%, cache hits 57%, output 9%; off-peak stays at half of peak. The larger move is on V4-Pro: from 04:00 UTC on September 14, DeepSeek routes every deepseek-v4-pro request to V4.1-Flash at V4.1-Flash rates “until V4.1-Pro launches”. The pricing page still lists V4-Pro at $1.32 / $3.96, and nothing is billed at that rate. Until a new Pro appears, DeepSeek’s whole API costs $0.30 / $1.20 at peak.

New pricing effective 04:00 UTC, September 10 (DeepSeek’s release note). The page now defines peak as 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays; weekends are off-peak in full.

Effective

OpenAI’s GPT-6 Astra: the frontier slot doubles on input

Model evolutionFrontier slot
New modelPredecessorNewChange
gpt-6-astra (frontier)listed above gpt-5.6-solInput$5→$10+100%
Cached input$0.50→$1+100%
Output$30→$50+67%

GPT-6 Astra reached the API on September 3 as “our most capable model”, listed above the gpt-5.6 line and recommended as the place to start. Priced against gpt-5.6-sol at its list rate, input doubles, output rises by two thirds, and the cached-input rate doubles. Batch and flex stay at half of standard; past 272k input tokens the rate is $20 / $75. sol stays listed, now at a promotional rate (entry below). Since gpt-5 the frontier slot has gone from $1.25 / $10 to $10 / $50: 8x on input, 5x on output.

Effective

Claude Fable 5.1: headline unchanged, cache reads cut by three quarters

Model evolutionFrontier slot
New modelPredecessorNewChange
Claude Fable 5.1succeeds Claude Fable 5Input$10→$100%
Cache read$1→$0.25-75%
Output$50→$500%

Anthropic launched Claude Fable 5.1 on September 1 as the successor to Claude Fable 5, at the same $10 / $50. The change is on the cache: a cache hit costs 2.5% of the input rate, where every other Claude model charges 10%. For a workload that re-sends a long prefix on every call, that is a real cut the headline does not show. At a 90% cache-hit rate the effective input rate falls from $1.90 to $1.23 per million tokens. GPT-6 Astra, at the same list price, keeps a $1 cache read. Fable 5 itself, launched in June at $10 / $50, predates this log; Anthropic’s Opus line stays at $5 / $25.

Effective

gpt-5.6-sol drops to a promotional rate

Price changeOne model
ModelBeforeAfterChange
gpt-5.6-solstandard rateInput$5→$4-20%
Cached input$0.50→$0.40-20%
Output$30→$20-33%

OpenAI’s API changelog dates the cut August 21 and calls the rate promotional, “available at least through November 21, 2026”, without naming the rate that follows. Batch and flex move with it to $2 / $10; long context to $8 / $30. It is the second in-place cut on the gpt-5.6 line in three weeks (terra and luna, entry below), and it lands two weeks before a model at two and a half times its price is listed above it. This log caught it on September 7.

Effective

DeepSeek repriced up roughly 3x and introduced peak and off-peak rates

Price changeFull price list
ModelBeforeAfterChange
deepseek-v4-propeakInput$0.435→$1.32+203%
Output$0.87→$3.96+355%
off-peakInput$0.435→$0.66+52%
Output$0.87→$1.98+128%
deepseek-v4-flashpeakInput$0.14→$0.44+214%
Output$0.28→$1.32+371%
off-peakInput$0.14→$0.22+57%
Output$0.28→$0.66+136%

Announced and effective 16:00 UTC with the V4-Pro release. Peak hours are 01:00–04:00 and 06:00–10:00 UTC; off-peak is half of peak.

Effective

gpt-5.6-terra cut 20% and gpt-5.6-luna cut 80%, three weeks after launch

Price changeTwo models
ModelBeforeAfterChange
gpt-5.6-terraInput$2.50→$2-20%
Output$15→$12-20%
gpt-5.6-lunaInput$1→$0.20-80%
Output$6→$1.20-80%

OpenAI’s API changelog: “Starting July 30, GPT-5.6 Luna costs 80% less, while GPT-5.6 Terra costs 20% less.” The cheap slot returns to its predecessor’s input rate and sits 4% under it on output; the mid slot lists 20% under gpt-5.4. The same entry replaced Priority Processing with Fast mode, at twice the standard rate for sol.

Effective

OpenAI’s gpt-5.6 generation: the frontier holds, and the two slots below it launch at three to five times their predecessors

Model evolutionFull lineup
New modelPredecessorAt launchChange
sol (frontier)succeeds gpt-5.5Input$5→$50%
Output$30→$300%
terra (mid)succeeds gpt-5.4-miniInput$0.75→$2.50+233%
Output$4.50→$15+233%
luna (cheap)succeeds gpt-5.4-nanoInput$0.20→$1+400%
Output$1.25→$6+380%

Sol was previewed on June 26 and the family reached the API on July 9, three models where the prior generation offered five. The pairings above are OpenAI’s own: its model pages describe sol as corresponding to the unsuffixed tier of earlier GPT-5 families, terra to the mini tier, and luna to the nano tier. On that reading the frontier rate held and both slots beneath it launched well above what they replaced. Three weeks later OpenAI cut luna 80% and terra 20%, which is the entry above. This was a succession rather than a removal: gpt-5.4, its mini and nano variants and gpt-5.5 are all still sold at unchanged prices, on their own model pages rather than the main pricing table.

Launch rates for terra and luna are worked back from OpenAI’s July 30 changelog entry against the rates listed since; the July 9 changelog entry names the models but not their prices.

Corrected September 21, 2026. This entry first ran dated by our read date, August 21, with the mid slot 20% cheaper and the cheap slot’s output down 4%. Those were the rates after OpenAI’s July 30 cut, not at launch, and terra was paired against gpt-5.4 rather than the mini tier OpenAI’s own page names. The date, the launch rates and the pairing are all corrected here, and the July 30 cut now has its own entry.

Effective

Anthropic: the same text can cost up to a third more, at an unchanged headline price

Effective price changeClaude 4.7 and later

Claude Opus 4.7 shipped a new tokenizer on April 16, and Anthropic’s launch note says the same input can map to roughly 1.0 to 1.35 times as many tokens. The per-token price did not change, so for text at the top of that range the cost of a unit of real work rose by about a third, and for text at the bottom it did not move at all. Which end a workload sits at depends on its content, and Anthropic publishes no breakdown. It is a price change that appears on no pricing page and that a per-token comparison cannot see.

Corrected September 21, 2026. This entry first ran dated by our read date, August 21, and gave approximately 30% as Anthropic’s own figure. Anthropic published a range, 1.0 to 1.35 times, and the entry now quotes it. The effective date is the Claude Opus 4.7 launch.

Get new entries by email

One email per entry, nothing else.

Was this helpful?