LLM API price changes, tracked.

GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini

Deprecated on 14 October 2026 and scheduled for removal from the API on 14 January 2027, with gpt-6-luna named as the recommended replacement for all four. OpenAI says fine tuned versions follow the same shutdown schedule as their underlying base models.

So what
Three months, and one named replacement for models of very different sizes. GPT-6 Luna lists $0.10 input and $0.50 output per million tokens for short context, so the move is likely cheaper per token, but run your own evals rather than assuming parity, and plan to rebuild any fine tune on a new base model.

Gemini 4 Argon

Generally available on 13 October 2026 at $2 input and $10 output per million tokens, which the pricing page lists as running through 30 November 2026 and rising to $4 and $20 starting 1 December 2026. Context caching goes from $0.10 to $0.20, batch and Flex from $1 and $5 to $2 and $10, and Priority from $3.60 and $18 to $7.20 and $36. Output pricing includes thinking tokens, only implicit caching is supported, and the model is not on the free tier or the Google AI Pro plan.

input / 1M

$2.00 -> $4.00

output / 1M

$10.00 -> $20.00

So what
A doubling with a published date, on a model that launched level with Claude Sonnet 5.5 and GPT-6.1 Sol at $2 and $10. Cost your product at $4 and $20, and compare cost per successful task on your own prompts, because the December comparison is double the launch one.

Claude Haiku 5.5

Launched 7 October 2026 and priced by prompt length. For prompts up to 100,000 tokens it is $0.10 input and $0.50 output per million tokens, with cache reads at $0.01. For prompts over 100,000 tokens it is $0.50 and $2.50, with cache reads at $0.05. The pricing page says prompt length counts all input tokens including cache reads and writes, and each request is priced on its own. Claude Haiku 4.5 is $1 and $5. Haiku 5.5 uses the Claude 4.7 tokenizer, which the migration guide says produces approximately 30 percent more tokens for the same text than Haiku 4.5.

input / 1M

$1.00 -> $0.10

output / 1M

$5.00 -> $0.50

So what
A tenth of the Haiku 4.5 rate below the line and half of it above. With about 30 percent more tokens the saving is nearer 87 percent below the line and 35 percent above it, and a prompt of roughly 77,000 tokens by Haiku 4.5 counting lands at the line. Measure your prompt length distribution on the new tokenizer before you model a saving. The roster here lists Haiku 4.5 only, because its single rate per model cannot express a threshold.

Claude Sonnet 5.5 (cache reads)

On 7 October 2026 Anthropic lowered Claude Sonnet 5.5 prompt cache reads from $0.20 to $0.10 per million tokens, which its release notes describe as 0.05x the base input price instead of 0.1x. Cache writes and all other prices are unchanged. The pricing page lists Claude Opus 5.5 at the same 0.05x multiplier. Input and output stay at $2 and $10.

So what
Our Sonnet 5.5 entry of 2 October recorded cache reads at $0.20, which was right then and is stale now. On cached input Sonnet 5.5 now matches the $0.10 on GPT-6.1 Sol, so the comparison we made in issue 5 no longer favours OpenAI. On the five minute cache duration a prefix pays for itself after a single read.

Claude Sonnet 5.5

Released 28 September 2026 at $2 input and $10 output per million tokens, with cache reads at $0.20, against Claude Sonnet 4.5 at $3 and $15 with cache reads at $0.30. Sonnet 4.5 was deprecated on 30 September with retirement on Anthropic-operated platforms on 30 November 2026. The migration guide notes that Sonnet 5.5 uses Sonnet 5's tokenizer, and that against Sonnet 4.6, Sonnet 4.5 and Haiku 4.5 the same text produces about 30 percent more tokens.

input / 1M

$3.00 -> $2.00

output / 1M

$15.00 -> $10.00

So what
A third off the rate is not a third off the bill. At 0.667 times the rate and about 1.30 times the tokens, the same English text lands near 13 percent cheaper, and a 2000x1500 image costs about 2.5 times as many tokens because 5.5 uses the high resolution image tier. Coming from Sonnet 5 there is no token change, so the full third applies. This is the clearest case we have logged for comparing providers on measured token counts rather than on rate cards.

GPT-6.1 Sol

Released 29 September 2026 at $2 input, $0.10 cached input, $2.50 cache write and $10 output per million tokens for prompts up to 272K input tokens, against GPT-6 Astra at $10 input, $1 cached and $50 output. Context window is 1,050,000 tokens with 128,000 maximum output. OpenAI describes it as being for complex coding and professional work at a lower cost than GPT-6 Astra.

input / 1M

$2.00

output / 1M

$10.00

So what
A fifth of Astra's rate on both input and output, and half the cached input rate of GPT-6 Sol released a week earlier, which matters most if you resend a large fixed prompt. Two caveats: there is still no dated snapshot id for the GPT-6 family, only the bare alias, so you cannot pin a version, and input doubles above 272K input tokens while output rises by half.

GPT-6 Sol and GPT-6 Luna

Released 22 September 2026. For prompts up to 272K input tokens, GPT-6 Sol is listed at $2 input, $0.20 cached input and $10 output per million tokens, and GPT-6 Luna at $0.10, $0.01 and $0.50. Above 272K, Sol is $4, $0.40 and $15, and Luna is $0.20, $0.02 and $0.75. GPT-5.6 Sol and GPT-5.6 Luna remain listed at $4 and $20, and $0.20 and $1.20.

input / 1M

$2.00

output / 1M

$10.00

So what
Half the listed price of the 5.6 models sitting beside them. Note the asymmetry above 272K input tokens: input and cached input double, but output rises by half, not double, so a long-context budget built on a flat doubling is wrong. Nothing on OpenAI's pages describes GPT-6 pricing as promotional, unlike GPT-5.6 Sol.

Claude Opus 5.5

Released 22 September 2026 at $4 input and $20 output per million tokens, against Opus 5 at $5 and $25, which stays listed. Cache hits are priced at 0.05x base input rather than the usual 0.1x, putting a cache hit at $0.20 per million. Anthropic says it costs 40 percent less than Opus 5 on typical workloads at default settings, a larger figure than the 20 percent list cut because it also claims token efficiency gains.

input / 1M

$4.00

output / 1M

$20.00

So what
Not a drop-in swap. Three request shapes now return a 400 error: the thinking field in either its enabled or disabled form, tool_choice types any and tool, and the computer_20251124 computer use tool on the Claude API and Google Cloud, though that one keeps working on Amazon Bedrock. On cache-heavy agentic work the 0.05x multiplier matters more than the headline rate.

Gemini 3.6, 3.7 and 3.8 Flash

The Gemini Developer API pricing page lists input at $0.75 through 31 December 2026 and $1.50 from 1 January 2027, with output going from $3.75 to $7.50 per million tokens. The same date raises both Gemini 3.8 TTS models, both Gemini Robotics ER 2 previews, and context caching, which goes from $0.075 to $0.15 per million on 3.8 Flash. Gemini 3.5 Flash Lite at $0.30 and $2.50 has no scheduled increase, and neither does any Pro tier.

input / 1M

$0.75 -> $1.50

output / 1M

$3.75 -> $7.50

So what
A doubling with a published date is the most actionable kind of entry on this page, because you can still move. Build the 2027 budget on $1.50 and $7.50 and benchmark against Flash Lite, which carries no scheduled rise.

Gemini 3.7 and 3.8 Flash

Our roster described the $0.75 and $3.75 rates as an introductory rate through 31 December 2026. Google's pricing page does not use the words introductory or promotional anywhere. It lists two prices with a date between them. The note now reads as a scheduled increase.

So what
The distinction is not pedantry. An introductory rate implies a provider decision that could be extended, while a dated increase on a pricing page is a commitment you should plan around. We described Google as doing the first when it had done the second.

DeepSeek V4.1 Flash (was V4 Flash)

DeepSeek retired V4 Flash and now lists DeepSeek-V4.1-Flash under the API name deepseek-flash at $0.15 input and $0.60 output off-peak, $0.30 and $1.20 peak. The legacy name deepseek-v4-flash is still accepted, served by V4.1 Flash and billed at the new price. V4 Pro is unchanged and DeepSeek says it will keep serving it past 14 September.

input / 1M

$0.22 -> $0.15

output / 1M

$0.66 -> $0.60

So what
A cut you get without touching code: requests to the old model name already bill at the new price. Our weekly drift check caught this on 14 September because $0.22 vanished from the page. After last month, where we dismissed two real flags as noise, this one was verified against the rendered page before anything changed.

gpt-5.4-cyber

Deprecated on 11 September 2026 and removed from the API on 1 October 2026. OpenAI names gpt-5.6-cyber as the replacement.

So what
Twenty days from announcement to removal, much shorter than the months OpenAI usually gives. If you run security tooling on it, this is a two-week migration, and the kind of short notice worth a calendar alert on the deprecations page itself.

o3-mini replacement (our GPT-5.6 page)

Our GPT-5.6 page described Terra as the replacement for both o3-mini and o4-mini. OpenAI's deprecation table names gpt-5.6-sol for o3-mini and gpt-5.6-terra for o4-mini. Corrected.

So what
Terra is still the cheaper landing if it holds up in your evals, but it is not what OpenAI recommends for o3-mini, and Sol costs roughly four times o3-mini per token. We now state both the named mapping and the price trade-off, instead of blurring them into one recommendation.

o3-mini, o4-mini

Both shut down on 23 October 2026, announced 22 April 2026 alongside o1, o1-pro and the legacy GPT-4 snapshots. OpenAI points migrations at the GPT-5.6 family.

input / 1M

$1.10

output / 1M

$4.40

So what
Six weeks. These are reasoning models, so they tend to sit on the escalation path rather than the default one, which is exactly where a failure is least likely to be noticed in testing and most expensive in production. Grep for o3-mini and o4-mini today.

o3-mini, o4-mini (our listing)

We listed both as current with no caveat. They have had a published shutdown date since 22 April 2026. Our own weekly drift check flagged both on 5 September and we recorded it as a likely false positive.

So what
The tool was right and the human reading it was wrong. Both entries now carry the shutdown date. If you sized a budget here on o3-mini in the last month, you priced a model that stops answering on 23 October.

Whisper and the gpt-4o transcribe family

whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize were deprecated on 26 August 2026 and shut down on 26 February 2027. Replacements are gpt-transcribe and gpt-live-transcribe.

input / 1M

$0.01

So what
Whisper has one of the largest install bases of any OpenAI endpoint, and a lot of it is glue code written once and never revisited. Five months is generous by OpenAI standards; the risk is that nobody looks until it is two weeks out.

Videos API and Sora 2

The Videos API and the sora-2 and sora-2-pro models are removed on 24 September 2026, announced 24 March. The deprecation notice names no replacement.

So what
Seventeen days, and unlike every other entry on this page there is nowhere to migrate to within OpenAI. If you ship video generation on Sora, this is a vendor change, not a version bump.

Gemini 3.8 Flash, 3.7 Flash, 3.6 Flash

All three Flash tiers are listed at an introductory $0.75 input and $3.75 output per million tokens, marked as running through 31 December 2026.

input / 1M

$0.75

output / 1M

$3.75

So what
If you are sizing a 2027 budget on today's Flash rate you are sizing it on a promotional number with a published end date. Model the post-January figure now, or pin the workload to a tier whose price is not time-boxed.

Llama 3.3 70B, Llama 3.1 405B

Groq no longer publishes a per-token price for Llama 3.3 70B, the model page now says contact sales. Together still lists Llama 3.3 70B, at $1.04 input and $1.04 output, and no longer lists Llama 3.1 405B at all.

input / 1M

$0.59 -> $1.04

output / 1M

$0.79 -> $1.04

So what
Open-weight hosting prices are drifting out of public view, which makes them harder to budget than closed models rather than easier. If your cost model assumes a published Groq rate, it is now an assumption rather than a number. We show the Together rate because it is the one you can actually read.

Command R+

Command R+ 08-2024 is still $2.50 input and $10 output, but Cohere now lists the whole Command family under legacy models available to existing customers.

So what
The price has not moved, the availability has. If you are not already a Cohere customer this is not a model you can newly adopt at these rates, so drop it from any build-versus-buy comparison you are running today.

Sonar, Sonar Pro, Sonar Reasoning Pro

Token prices are unchanged (Sonar Pro at $3 and $15), but Perplexity also charges a per-request search fee of $5 to $14 per 1,000 requests depending on search context size, and Deep Research bills citation and reasoning tokens separately.

So what
Token pricing alone understates Sonar badly. At high context, 1,000 requests carries $14 of search fees before a single token is counted, which on short exchanges can exceed the token cost outright. Model this one per request, not per million tokens.

Grok 4.6

xAI now lists Grok 4.6 at $2 input and $6 output per million tokens. Grok 4, which we had at $3 and $15, is no longer on the models page.

input / 1M

$3.00 -> $2.00

output / 1M

$15.00 -> $6.00

So what
The current top Grok is a third of the output price of the one it replaced. Rates double on prompts of 200k tokens or more, so long-context workloads should budget at the higher tier rather than the headline.

Mistral Large

Mistral lists Large at $0.50 input and $1.50 output per million tokens. We were carrying Mistral Large 2 at $2 and $6.

input / 1M

$2.00 -> $0.50

output / 1M

$6.00 -> $1.50

So what
Four times cheaper on input than the figure we published. Mistral is now one of the cheapest ways to run a capable open-weight-adjacent model, and worth re-testing if you priced it out earlier.

DeepSeek V4 Pro and V4 Flash

DeepSeek now lists V4 Pro at $0.66 input and $1.98 output, and V4 Flash at $0.22 and $0.66, both off-peak. V3 and Reasoner are gone from the pricing page.

input / 1M

$0.66

output / 1M

$1.98

So what
DeepSeek bills on a peak and off-peak clock: peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, and the rate doubles. A batch job you can shift outside those windows literally costs half as much.

Gemini 2.5 Pro, 2.5 Flash, 2.5 Flash-Lite

The entire Gemini 2.5 line now appears under "legacy models" on Google's pricing page. Prices are unchanged at $1.25/$10, $0.30/$2.50 and $0.10/$0.40.

So what
Legacy listing is the step before a retirement date. Nothing breaks today, but 2.5 is no longer where Google puts new capability, and a migration you plan now is cheaper than one you do on a deadline.

Claude Opus 5

Opus 5 is listed as current at $5 input and $25 output per million tokens. Opus 4.8, at $15 and $75, has moved to the legacy list.

input / 1M

$15.00 -> $5.00

output / 1M

$75.00 -> $25.00

So what
The current Opus tier costs a third of the one it replaces. If your routing config still escalates to Opus 4.8 you are paying 3x for the older model, and that is a one-line change.

GPT-6 Astra

GPT-6 Astra is listed at $10 input and $50 output per million tokens, above GPT-5.6 Sol at $4 and $20.

input / 1M

$10.00

output / 1M

$50.00

So what
Astra is 2.5x Sol on both sides. Treat it the way you would treat any top tier: an escalation path for the calls that need it, not the default in your client config.

GPT-5

Our roster carried GPT-5 at $5 input and $15 output. OpenAI lists $1.25 and $10. Corrected.

input / 1M

$5.00 -> $1.25

output / 1M

$15.00 -> $10.00

So what
Our calculator overstated GPT-5 input cost by 4x for anyone who modelled a bill on it before 5 September 2026. If you sized a budget here and ruled GPT-5 out on price, run it again.

Claude Sonnet 5

Our roster carried Sonnet 5 at $3 input and $15 output. Anthropic lists $2 and $10. Corrected.

input / 1M

$3.00 -> $2.00

output / 1M

$15.00 -> $10.00

So what
A third cheaper than we were showing, on both sides. Same re-run advice: if Sonnet 5 lost a cost comparison here before 5 September 2026, the comparison was wrong.