LLM API price change log
Prices verified August 30, 2026 ยท ํ๊ตญ์ด ยท ๆฅๆฌ่ช
Vendor pricing pages tell you what a model costs today. They almost never tell you what it cost last month, or that the number moved at all. This page is the missing half: every LLM API price change we can verify against the vendor's own page, with the date it took effect, the old rate, the new rate, and what it does to a real monthly bill.
We add an entry when a price actually moves โ not on a publishing schedule. A quiet month means nothing changed.
August 2026 produced three changes across three vendors, in three different directions: DeepSeek raised prices, OpenAI cut them, and Anthropic withdrew an increase it had already announced. If you set a budget at the start of the month, all three moved under it.
OpenAI cut GPT-5.6 Sol
21 August 2026 ยท decrease
$5 / $30 became $4 / $20 โ the flagship got cheaper
| Rate (per 1M tokens) | Before | After | Change |
|---|
| Input | $5.00 | $4.00 | −20% |
| Cached input | $0.50 | $0.40 | −20% |
| Output | $30.00 | $20.00 | −33% |
Read the wording carefully. OpenAI says the rate is “available at least through November 21, 2026”. “At least through” is a guaranteed floor, not an expiry date. On 21 November the price may rise, may be extended, or may fall further. Plenty of write-ups have already turned that phrase into a deadline. It isn't one.
Terra, Luna and the GPT-4o tier were unchanged. Worth knowing alongside it: on GPT-5.6 models a request whose input exceeds 272K tokens reprices the entire call at 2× input and 1.5× output โ not just the tokens above the line.
Source: developers.openai.com/api/docs/pricing, Standard tier · short context. Read on screen 30 August 2026. Several aggregator sites still list the pre-cut $5 / $30.
DeepSeek repriced its entire API
16 August 2026 ยท increase
Output roughly tripled, and the day is now split into peak and off-peak
| Rate (per 1M tokens) | Before | Off-peak | Peak | Change |
|---|
| V4 Pro input | $0.435 | $0.66 | $1.32 | +52% |
| V4 Pro output | $0.87 | $1.98 | $3.96 | +128% |
| V4 Pro cache hit | $0.003625 | $0.022 | $0.044 | +507% |
| V4 Flash input | $0.14 | $0.22 | $0.44 | +57% |
| V4 Flash output | $0.28 | $0.66 | $1.32 | +136% |
| V4 Flash cache hit | $0.0028 | $0.007 | $0.014 | +150% |
Peak is 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. Everything else is off-peak, at exactly half the peak rate. For teams in the Americas and Europe that window falls overnight, so most interactive traffic bills off-peak โ but scheduled jobs and Asian-timezone traffic can sit entirely inside it. Moving a nightly batch by two hours is worth half its bill.
Context is 1M tokens, not the 128K many tables still show, with a 384K maximum output.
What it does to a real bill. Support chatbot, 3,500 input + 400 output tokens per request, 30,000 requests/month, 80% cache hit rate:
V4 Pro before: $19.88/month → after, off-peak: $39.47/month
V4 Flash before: $6.54/month → after, off-peak: $13.13/month
Source: api-docs.deepseek.com/quick_start/pricing, cache-miss rates. Read on screen 30 August 2026.
Why the DeepSeek change reorders the ranking
DeepSeek's real advantage was never the sticker price โ it was the cache ratio. A cache hit cost 0.83% of a cache miss, against roughly 10% at OpenAI, Anthropic and Google. That is what let V4 Pro overtake models with a lower list price once enough of your input was cached.
After the repricing a V4 Pro cache hit costs 3.33% of a miss โ still the deepest cache discount of any major vendor, but no longer deep enough to pay for a 128% higher output rate.
| Chatbot · 3,500 in / 400 out · 30k req/mo | 0% cached | 50% | 80% | 95% |
|---|
| GPT-5.6 Luna | $35.40 | $25.95 | $20.28 | $17.45 |
| DeepSeek V4 Pro before 16 Aug | $56.12 | $33.47 | $19.88 | $13.09 |
| DeepSeek V4 Pro now, off-peak | $93.06 | $59.56 | $39.47 | $29.42 |
| DeepSeek V4 Flash now, off-peak | $31.02 | $19.84 | $13.13 | $9.77 |
Read rows two and three against the first. Before the change, V4 Pro overtook GPT-5.6 Luna at a 78.5% cache hit rate. After it, V4 Pro is more expensive than Luna at every hit rate from 0% to 100% โ the crossover is gone entirely.
V4 Flash still wins at every hit rate, and by a wide margin. DeepSeek remains the cheapest way to serve this workload; the model you want is simply Flash now, not Pro. If you chose V4 Pro specifically for its cache economics, that reason no longer exists.
One crossover does survive at current prices: V4 Pro passes Gemini 3 Flash at a 23.10% hit rate, where both land on $77.59/month. Gemini 3 Flash has the cheaper input ($0.50 vs $0.66) and still loses once a fifth of the prompt is cached.
Our own calculation from the rates above. Reproduce it with your token mix in the calculator.
Claude Sonnet 5's increase was withdrawn
10 August 2026 ยท increase cancelled
$2 / $10 is the standard price โ the September 1 rise to $3 / $15 will not happen
Sonnet 5 launched with $2 input / $10 output described as introductory pricing through 31 August 2026, with a rise to $3 / $15 scheduled for 1 September. Anthropic withdrew that increase and made the launch price permanent.
Anthropic's documentation states: “The $2/$10 per million input/output token pricing for Claude Sonnet 5, announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur.”
If you budgeted a 50% increase into your September forecast, take it out. Plenty of comparison sites โ and, until late August, this one โ still carried the old “promotional until 31 August” wording.
Source: platform.claude.com/docs/en/about-claude/pricing. Announcement dated 10 August 2026. Read on screen 30 August 2026.
Correction: Gemini 3.6 Flash was never cut. That was us misreading a tab.
30 August 2026 ยท our error, corrected
We published $0.75 / $3.75 as a price cut. It is the Batch tab.
For part of 30 August this site showed Gemini 3.6 Flash at $0.75 input / $3.75 output and described it as a 50% cut effective 13 August, reverting 1 January 2027. None of that happened. The standard rate is and was $1.50 / $7.50.
| Gemini 3.6 Flash | Input | Cached | Output |
|---|
| Standard tab โ correct | $1.50 | $0.15 | $7.50 |
| Batch tab | $0.75 | $0.075 | $3.75 |
| What we published | $0.75 | $0.075 | $3.75 |
We are logging our own mistake because the mechanism behind it will catch other people, and because a “Gemini 3.7 Flash” briefly appeared in our tables that does not exist in Google's model list at all.
Why it slipped through. OpenAI stacks its tiers vertically โ scroll far enough and you meet a heading that says Batch. Google puts them in tabs: Standard / Batch / Flex / Priority swap the table in place, and scrolling never reveals that the other tabs exist. A summary tool reading the page has no idea which tab was active.
Why it was convincing. Batch is exactly 50% off at most vendors, so a Batch table read as standard looks precisely like a half-price cut โ and a tidy story assembles itself around it, expiry date included.
The rule we now follow: a price is not confirmed until we have read the table's own heading on screen. An AI API price table has at least three axes โ tier (Standard / Batch / Flex / Priority) × context length (short / long) × cache state. Any figure quoted without naming all three is ambiguous. And “it halved” should raise suspicion of a Batch tab before it raises hope of a price cut.
Corrected against ai.google.dev/gemini-api/docs/pricing, Standard tab, read on screen 30 August 2026. All other Gemini models were re-checked on the Standard tab at the same time and were already correct.
How to use this page
Re-check monthly, not at procurement. Three vendors moved inside three weeks, and none announced it through a channel most teams monitor.
Watch the cache rate, not just the input rate. In the same DeepSeek change, input rose 52% while cache-hit rose 507% โ roughly ten times as far. If your prompt has a large stable prefix, the cache rate is deciding most of your bill.
Re-run the comparison; don't reason from the direction. “DeepSeek got more expensive but it's still the cheapest” is intuitive, and for V4 Pro against GPT-5.6 Luna it is wrong.
Check which table you are reading. The entry above is what happens when you don't.
→ Re-price your workload at current rates
Enter your token mix once โ every model ranked, with peak/off-peak, caching and per-vendor batch discounts applied.
Advertisement
Current prices by vendor
Frequently asked questions
Did Claude Sonnet 5 prices go up on September 1, 2026?
No. On 10 August 2026 Anthropic confirmed $2 / $10 is now the standard price and the scheduled rise to $3 / $15 will not occur.
When are DeepSeek's peak hours?
01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. Every other hour bills at the off-peak rate, exactly half the peak rate.
Does the GPT-5.6 Sol price expire on 21 November 2026?
Not necessarily. OpenAI's wording is “available at least through November 21, 2026” โ a guaranteed floor, not an expiry. The price may rise, be extended, or fall further on that date.
Was Gemini 3.6 Flash cut to $0.75 / $3.75?
No. $1.50 / $7.50 is the standard rate. $0.75 / $3.75 is the Batch tab on Google's pricing page. We published that error ourselves and corrected it the same day.
Every figure here was read on the vendor's own pricing page, on the Standard tier at short context โ not Batch, Flex or Priority. Each entry names the source and the date we read it. Vendors change prices without notice and regional rates vary; confirm against the vendor's page before committing budget. Cost figures are our own calculation from the listed rates.