What launched, and which tier each model is in
Three launches in seventeen days, from three providers, with three different stories attached.
- 26 October: OpenAI released GPT-6.1 Terra at $1 input and $5 output per million tokens. OpenAI describes it as 50% cheaper than GPT-6.1 Sol, which lists at $2 and $10, and the arithmetic matches.
- 9 November: Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 at $8 input and $40 output, down 20% from Fable 5.1's $10 and $50. A fast mode costs $16 and $80.
- 12 November: Google launched Gemini 4 Hydrogen at $1.50 input and $7.50 output, calling it "our daily driver for agentic workflows".
Comparing these headline prices directly would be a mistake, because they are not the same kind of product. Terra and Hydrogen are each provider's fast, cheap workhorse. Fable 5.2 is Anthropic's top general-availability model, and Anthropic still recommends Claude Opus 5.5, at $4 and $20, as the default for most work, positioning Fable for long-horizon agentic tasks or jobs where Opus 5.5 falls short. The fair comparisons are Terra against Hydrogen, and Fable 5.2 against the other providers' top models.
Where each sits
Google is explicit about the split in its own family: Hydrogen is "built for rapid execution", while Gemini 4 Carbon and Argon are the frontier models that "excel at deep, long-horizon reasoning". It also claims Hydrogen scores 51 on the Artificial Analysis Intelligence Index against Argon's 50 at default settings. That is a vendor claim about its own models, but if it holds up, it is the most commercially important sentence in the launch.
Mythos 5.2 needs a separate note. It shares Fable 5.2's specifications and pricing, but access is limited to organisations verified through Anthropic's programmes. Most buyers can ignore it for budgeting purposes.
Where you can buy them
Fable 5.2 is available on the Anthropic API, in Claude apps and Claude Code, and through Google Cloud, Microsoft Foundry and Anthropic's platform on AWS. Hydrogen is on the Gemini API and AI Studio, in Google's enterprise products and in the consumer Gemini app. Terra is on the OpenAI API. If your procurement runs through one cloud's marketplace, availability there may narrow the choice before price does; our Google Vertex AI versus OpenAI API comparison covers the platform side.
The prices, side by side
All figures are list prices per million tokens on each provider's standard tier, taken from their own pricing pages or launch material on 12 November 2026.
| Model | Input | Cached input | Output | Notes |
|---|---|---|---|---|
| GPT-6.1 Terra | $1.00 | $0.05 | $5.00 | $2.00 and $7.50 above 272K input tokens |
| Gemini 4 Hydrogen | $1.50 | $0.075 | $7.50 | Flat across 1M context; free tier available |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | $4.00 and $15.00 above 272K |
| Gemini 4 Carbon or Argon, to 30 Nov | $2.00 | $0.10 | $10.00 | Introductory; not on free tier |
| Gemini 4 Carbon or Argon, from 1 Dec | $4.00 | $0.20 | $20.00 | As listed on Google's pricing page |
| Claude Opus 5.5 | $4.00 | Not covered here | $20.00 | Anthropic's recommended default |
| Claude Fable 5.2 | $8.00 | Not covered here | $40.00 | Fast mode $16 and $80 |
We have not listed Anthropic's cached input rates because we could not confirm them against Anthropic's own page for this article; check the Anthropic API pricing page before modelling a cache-heavy workload. Our OpenAI API pricing page carries the full OpenAI table, including the modes below.
Batch and other modes
Both Google and OpenAI halve prices for asynchronous work. Terra on batch or flex is $0.50 and $2.50; Hydrogen on batch is $0.75 and $3.75. OpenAI's fast mode doubles Terra to $2 and $10. Anyone running overnight jobs, evaluations or bulk classification at standard rates is paying twice what they need to, whichever provider they use.
Thinking tokens are output tokens
Google's pricing page states that output is priced "including thinking tokens" and that "every token the model processes is counted and billed". The same principle applies across providers that bill reasoning: a model that thinks longer writes more output tokens, and output is the expensive side of every row above, at five times the input price. Hydrogen has four thinking levels, from low to a new xhigh; Fable 5.2 has its own effort settings. The effort setting can move a bill more than the choice of model.
The free tier comes with a condition
Hydrogen's input, output and caching are free of charge on Google's free API tier, and Carbon and Argon are not available there at all. Google's pricing table also marks free-tier data as "used to improve our products", against "no" for the paid tier. Free is fine for prototypes and public data. It is not the tier for customer records.
Three workloads, worked through
Prices per million tokens only mean something against a workload. Here are three, with the arithmetic shown so you can substitute your own volumes. All use standard-tier list prices and no caching unless stated.
Workload one: 10,000 support conversations a month
Each conversation sends 6,000 input tokens (system prompt, help-centre extracts and the customer's thread) and returns 600 output tokens. That is 60 million input and 6 million output tokens a month.
| Model | Input | Output | Monthly |
|---|---|---|---|
| GPT-6.1 Terra | 60 x $1.00 = $60 | 6 x $5.00 = $30 | $90 |
| Gemini 4 Hydrogen | 60 x $1.50 = $90 | 6 x $7.50 = $45 | $135 |
| GPT-6.1 Sol | 60 x $2.00 = $120 | 6 x $10.00 = $60 | $180 |
| Gemini 4 Carbon, to 30 Nov | 60 x $2.00 = $120 | 6 x $10.00 = $60 | $180 |
| Gemini 4 Carbon, from 1 Dec | 60 x $4.00 = $240 | 6 x $20.00 = $120 | $360 |
| Claude Opus 5.5 | 60 x $4.00 = $240 | 6 x $20.00 = $120 | $360 |
| Claude Fable 5.2 | 60 x $8.00 = $480 | 6 x $40.00 = $240 | $720 |
Two things stand out. First, at this volume every option is cheap next to the people it supports: the gap between the cheapest and dearest is $630 a month. Second, caching changes the picture more than model choice. If 4,000 of the 6,000 input tokens are a stable prefix, Terra's input falls to 40 x $0.05 + 20 x $1.00 = $22, and the month costs $52. Hydrogen falls to 40 x $0.075 + 20 x $1.50 = $33, or $78 a month.
If you buy AI support through a helpdesk rather than building it, the model price is hidden inside a per-resolution or per-seat fee; Zendesk and Intercom both price this way, and our analysis of per-resolution AI agent pricing shows how far that fee sits above the underlying token cost.
Workload two: a coding agent, one developer, one day
Fifty agent tasks a day, each re-sending context across its turns for a total of 300,000 input tokens per task, with every individual call under 272,000, and 20,000 output tokens per task. That is 15 million input and 1 million output tokens a day.
| Model | Arithmetic | Per day | Per month, 21 days |
|---|---|---|---|
| GPT-6.1 Terra | 15 x $1 + 1 x $5 | $20 | $420 |
| Gemini 4 Hydrogen | 15 x $1.50 + 1 x $7.50 | $30 | $630 |
| Gemini 4 Carbon, from 1 Dec | 15 x $4 + 1 x $20 | $80 | $1,680 |
| Claude Opus 5.5 | 15 x $4 + 1 x $20 | $80 | $1,680 |
| Claude Fable 5.2 | 15 x $8 + 1 x $40 | $160 | $3,360 |
| Claude Fable 5.2, fast mode | 15 x $16 + 1 x $80 | $320 | $6,720 |
Agents re-send most of their context on every turn, so in practice most of that input is cached. With 80% of input cached, Terra's day becomes 12 x $0.05 + 3 x $1 + $5 = $8.60, and Hydrogen's becomes 12 x $0.075 + 3 x $1.50 + $7.50 = $12.90. Anthropic also discounts cached reads, which would narrow the Fable figures, but see the note above on rates.
Seat-priced tools such as Cursor and Claude Code absorb or meter this differently; our Cursor versus GitHub Copilot comparison covers how the subscription products handle heavy agent use.
Workload three: long documents across the 272,000 line
A thousand contract reviews a month, each sending 300,000 input tokens and returning 5,000 output tokens. That crosses OpenAI's 272,000 token threshold; Google lists no long-context tier for these models.
| Model | Arithmetic | Monthly |
|---|---|---|
| GPT-6.1 Terra, 300K per request | 300 x $2.00 + 5 x $7.50 | $637.50 |
| Gemini 4 Hydrogen, 300K per request | 300 x $1.50 + 5 x $7.50 | $487.50 |
| GPT-6.1 Terra, trimmed to 260K | 260 x $1.00 + 5 x $5.00 | $285.00 |
| Gemini 4 Hydrogen, trimmed to 260K | 260 x $1.50 + 5 x $7.50 | $427.50 |
The cheaper model per token becomes the dearer one once requests cross the threshold, and becomes the cheaper one again if you trim the context by 40,000 tokens. Which provider wins this workload depends on a retrieval setting, not on a price list.
Why price per token is not cost per task
Anthropic's own launch framing makes this point, and the first independent numbers make it more sharply than Anthropic did.
Anthropic's claim
Anthropic argues Fable 5.2 is cheaper per task than its headline price suggests: at high effort it scores 45.3% on its FrontierCode 2.0 coding benchmark, roughly matching Opus 5.5's best, at about half the cost. That is a claim about a benchmark Anthropic runs, comparing two of its own models, and it may well be true for that task set.
What the first independent analysis found
An analysis published on 10 November by TokenCost, using Artificial Analysis Intelligence Index data, compared Fable 5.2 with Fable 5.1 at high effort, the API default:
| Measure, high effort | Fable 5.1 | Fable 5.2 | Change |
|---|---|---|---|
| List price, input and output | $10 and $50 | $8 and $40 | Down 20% |
| Output tokens per task | 66.0K | 97.8K | Up 48% |
| Cost per task | $7.54 | $9.08 | Up 20% |
The arithmetic is not mysterious. A 20% lower price multiplied by 48% more output gives 0.8 x 1.48 = 1.18, so roughly 18% more per task from output alone, before any change in input. The same analysis reports that at maximum effort the cost per task roughly doubled, and that at medium effort, the default in Claude apps and Claude Code, a task averaged $3.80, about 37% below Opus 5.5 at its highest effort for nearly the same score. The API defaults to high, so an integration that never sets effort pays about 2.4 times what it would on medium: $9.08 against $3.80.
This is not evidence that Fable 5.2 is poor value; it scores higher. It is evidence that a price cut tells you the rate, not the bill. The 20% in the headline and the 20% in the per-task cost point in opposite directions.
Google is making the same kind of claim
Google says early partner tests of Hydrogen showed 12% to 23% lower token use. If that holds on your workload, Hydrogen's effective cost is lower than its list price suggests. If it does not, it is not. Either way, it is the vendor's measurement, and the only one that counts for your budget is yours.
How to measure it yourself
Take a sample of a few hundred real tasks, run them through each candidate at the effort setting you would actually use, and record input tokens, output tokens, cached tokens and whether the task succeeded. Divide total cost by successful tasks. That number, cost per successful task, is the only one that compares a $1 model with an $8 one honestly. A router such as OpenRouter makes running the same sample across providers easier; see OpenAI API versus OpenRouter for the trade-offs.
Two cliffs to plan around
The 272,000 token threshold
OpenAI's pricing page defines short context as up to 272,000 input tokens and long context as anything above, and prices Terra's long-context requests at $2 input, $0.10 cached and $7.50 output: double the input and 1.5 times the output. When OpenAI introduced the same threshold on GPT-6 Astra, the higher rate applied to the whole request rather than the tokens above the line, which we worked through in our GPT-6 Astra pricing analysis. Terra's pricing page lists the long-context rate as a tier without restating that rule, so treat the whole-request behaviour as the working assumption and confirm it on your first invoice.
The practical fix is unchanged: log input tokens per request, look at the distribution rather than the average, and cap context assembly below 272,000 if a meaningful share of traffic sits near it. Workload three shows the payoff: trimming 40,000 tokens more than halved Terra's bill.
The 30 November introductory price
Gemini 4 Carbon and Argon are on introductory pricing of $2 input and $10 output "through November 30, 2026". Google's pricing page lists $4 and $20 from 1 December 2026, and the doubling applies across the board: cached input from $0.10 to $0.20, batch from $1 and $5 to $2 and $10. Any evaluation run on Carbon this month is being costed at half the price production will pay.
OpenAI has a date in the same week. Its pricing page states that GPT-5.6 Sol's promotional pricing runs until 30 November 2026. Two providers' promotional rates expiring together is a reason to re-run any cost model built in the last month, and it is the same pattern we tracked in this year's expiring promotional prices.
What to do before 1 December
Reprice every evaluation at post-promotion rates. If a Carbon pilot looked affordable at $2 and $10, rerun the numbers at $4 and $20 before signing anything. If it still looks affordable, good. If it only worked at the introductory rate, it did not work.
How to choose
There is no overall winner here, and anyone declaring one from list prices is skipping the step that matters. What the numbers do support:
- High-volume, short-context work such as classification, extraction and support replies: Terra is the cheapest per token at $1 and $5, with Hydrogen 50% dearer on both rates. Which is cheaper per successful task depends on token use and quality on your data.
- Very long prompts above 272,000 tokens: Hydrogen's flat $1.50 input undercuts Terra's $2 long-context rate. Trimming below the threshold reverses it.
- Prototypes and public data: Hydrogen's free tier costs nothing, with the condition that free-tier data may be used to improve Google's products.
- The hardest agentic and coding tasks: Fable 5.2 is the expensive option, and Anthropic itself recommends Opus 5.5 for most work. Pay for Fable where a cheaper model demonstrably fails, and measure cost per successful task, not per token.
- Carbon or Argon: budget at $4 and $20, not $2 and $10.
Do not lock in on a three-week snapshot
Three of the price lists in this article were set in the last seventeen days, and two more prices change on 30 November. Contracts with committed spend should be sized on measured usage and allow movement between models within a provider. Keeping the application able to switch providers, through an abstraction layer or a router, is cheap insurance against a market that reprices this often. Our OpenAI API alternatives and Anthropic API versus OpenAI API pages are a starting point, and the AI category lists the rest. For people choosing a chat product rather than an API, ChatGPT versus Claude covers the subscription side.
Questions people ask
- How much does Gemini 4 Hydrogen cost?
$1.50 per million input tokens and $7.50 per million output tokens on Google's paid tier, with cached input at $0.075 and batch at half price. Pricing is flat across the 1 million token context window. It is also free of charge on Google's free API tier, where data may be used to improve Google's products.
- How much does GPT-6.1 Terra cost?
$1 per million input tokens, $0.05 for cached input and $5 per million output tokens. Requests over 272,000 input tokens are billed at $2, $0.10 and $7.50. Batch and flex halve the standard rates, and fast mode doubles them.
- Is Claude Fable 5.2 cheaper than Fable 5.1?
Per token, yes: $8 and $40 per million tokens against $10 and $50, a 20% cut. Per task, not necessarily. An analysis of Artificial Analysis data found Fable 5.2 wrote about 48% more output tokens at high effort, and cost per task rose from $7.54 to $9.08.
- Will Gemini 4 Carbon and Argon get more expensive?
Yes. Both are $2 input and $10 output per million tokens through 30 November 2026. Google's pricing page lists $4 and $20 from 1 December 2026, with cached and batch rates doubling too.
- What is the 272K token threshold on OpenAI models?
OpenAI prices requests with more than 272,000 input tokens at long-context rates. For GPT-6.1 Terra that is $2 input and $7.50 output instead of $1 and $5. On GPT-6 Astra the higher rate applied to the whole request, not only the tokens above the threshold. See our GPT-6 Astra pricing analysis.
- Which AI model is cheapest in November 2026?
Of the models covered here, GPT-6.1 Terra has the lowest list price at $1 and $5 per million tokens, and Gemini 4 Hydrogen is the only one with a free API tier. For prompts above 272,000 tokens Hydrogen is cheaper than Terra. The cheapest model for a given job is the one with the lowest cost per successful task, which only your own testing can show.
Sources
Every price, limit and date above was checked against these pages on the day shown. Where a figure has since moved, the vendor’s own page is the authority and this one is a snapshot.
- 1Introducing Gemini 4 Hydrogen · GoogleLaunch date, $1.50 and $7.50 pricing, flat pricing across the context window, positioning and Google's token-use and benchmark claims.Checked
- 2Gemini Developer API pricing · GoogleHydrogen, Carbon and Argon rates, the 30 November introductory price and the $4 and $20 rate from 1 December, free-tier terms and batch prices.Checked
- 3API pricing · OpenAIGPT-6.1 Terra and Sol rates, the 272K long-context tier, batch and fast modes, and GPT-5.6 Sol's promotion ending 30 November.Checked
- 4Introducing Claude Fable 5.2 and Claude Mythos 5.2 · AnthropicThe 9 November launch, $8 and $40 pricing, fast mode, availability and Mythos 5.2 access limits.Checked
- 5Anthropic launches Claude Fable 5.2 and Mythos 5.2 with a 20% price cut as its IPO nears · MLQSecondary report of Anthropic's pricing, platforms, the Opus 5.5 recommendation and the per-task cost claim.Checked
- 6Claude Fable 5.2 Cost Per Task: Why Effort Decides · TokenCostIndependent analysis of Artificial Analysis data: output tokens and cost per task for Fable 5.2 against 5.1.Checked