GPT-6.2 Astra and GPT-6.2 Sol are two models that OpenAI released on November 17, 2026 (US time). Astra is the top model for the hardest work, and Sol is the main model for everyday work. They succeed GPT-6 Astra and GPT-6.1 Sol, respectively. For API pricing, Astra costs $8 input and $40 output per million tokens, which is 20% less than GPT-6 Astra. Sol costs the same as GPT-6.1 Sol, at $2 input and $10 output. In OpenAI's measurements, Astra beats the previous Astra on many tests at around half the cost, and Sol's scores rose on many tests at the same price. However, Claude Fable 5.2 beats it on some tests, and the API changes how mid-conversation rewrites are handled, so check a few things when you switch. This article summarizes pricing, performance, where the models are available, and how to choose between Astra and Sol, based on the values in OpenAI's announcement charts, the model pages, and the migration guide (checked November 19, 2026).
GPT-6.2 in three lines
Astra
Top model, 20% cheaper
$8 input and $40 output. It beats the previous Astra on all 10 tests compared in this article, and on many of them the cost per task fell to about half.
Sol
Smarter at the same price
Still $2 input and $10 output. On the FrontierCode test of real code changes, it is 0.4 points behind Astra, and it costs about one-fifth as much.
Note
How API conversations are built
Rewriting part of a conversation discards the earlier reasoning. Put shared instructions at the start of the input, not in instructions.
Checked: November 19, 2026. Sources are OpenAI's announcement, the GPT-6.2 Astra and GPT-6.2 Sol pages, and Using GPT-6.2. The performance figures are the values OpenAI included in its announcement charts.
1. The positioning of the two GPT-6.2 models
OpenAI's API model list opens with the guidance "If you're unsure, use GPT-6.2 Astra; for a balance, GPT-6.2 Sol; for low cost, GPT-6 Luna." The main models listed are GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra, and GPT-6 Luna. On its model pages, OpenAI describes Astra as "the most capable model, for taking the hardest projects from start to finish," and Sol as "the go-to model for everyday work, from coding to large-scale agents."
| Model | Input / output (per 1M tokens) | Knowledge cutoff | Previous model |
|---|---|---|---|
| GPT-6.2 Astra | $8.00 / $40.00 | May 18, 2026 | GPT-6 Astra ($10 / $50, knowledge cutoff April 30, 2026) |
| GPT-6.2 Sol | $2.00 / $10.00 | June 1, 2026 | GPT-6.1 Sol ($2 / $10, knowledge cutoff April 30, 2026) |
| GPT-6.1 Terra | $1.00 / $5.00 | June 1, 2026 | — |
| GPT-6 Luna | $0.10 / $0.50 | May 18, 2026 | — |
Source: OpenAI's Models page and each model's page (checked November 19, 2026). Both GPT-6.2 models have a context window of about 1.05 million tokens (up to 922K for input), a maximum output of 128K tokens, and accept text and image input. Terra and Luna did not change this time.
GPT-6 Astra and GPT-6.1 Sol remain available in the API, but they are no longer on the recommended list. OpenAI has not announced an end date for them, and they are not listed for retirement on the Deprecations page (as of November 19). Both new models also cannot set reasoning strength to "none (no reasoning)." For fast responses without reasoning, use GPT-6.1 Terra or GPT-6 Luna, which both accept "none." For details on the mid-tier Terra, see What is GPT-6.1 Terra?
2. Pricing: Astra gets a price cut, Sol stays the same
GPT-6.2 Astra
GPT-6.2 Sol
US dollars per 1 million tokens. The strikethrough prices are GPT-6 Astra's. Source: OpenAI's announcement and each model's page (checked November 19, 2026).
Cache writes cost 1.25 times the input price for 30-minute retention and 2.5 times for 12-hour retention. These multipliers are the same for both models.
| Item | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|
| Cache writes | $10 (30-minute retention) · $20 (12-hour retention) | $2.50 (30-minute retention) · $5 (12-hour retention) |
| Input over 272K tokens | Input and cache for the entire request are 2 times the normal price, and output is 1.5 times | |
| Batch | Half price | |
| Fast | 2 times | |
| Ultrafast | 3 times until February 28, 2027, and 6 times from March 1 (US Pacific time) | |
| Time-windowed background processing (beta) | 30% off for work that can be returned within 2 hours, 40% off if it can be returned within 8 hours (Responses API) | |
Sources: the GPT-6.2 Astra and GPT-6.2 Sol pages, and the API changelog (entry for November 17, 2026). Ultrafast is only available when data is stored in the US (from the migration guide).
According to OpenAI, Astra also uses fewer output tokens for the same work. As a result, the cost per task falls by more than the price cut alone, and in many tests it dropped to about half of GPT-6 Astra's cost, as the next chapter shows. Comparing prices with Claude, Astra matches Claude Fable 5.2 ($8 input and $40 output), Sol matches Claude Sonnet 5.5 ($2 input and $10 output), and Claude Opus 5.5 costs $4 input and $20 output (Anthropic's Pricing, checked November 19, 2026). Gemini 4 Carbon costs $2 input and $10 output until November 30, 2026, and $4 input and $20 output from December 1 (Google's Pricing, checked the same day). However, the number of tokens a task uses varies a lot between models, so comparing the cost per task is more reliable than comparing unit prices.
3. Performance ①: compared with the previous models
OpenAI's announcement includes 16 charts. Most show each model's scores and cost per task (some show output token counts) at each reasoning level. They break down into 10 tests of work, coding, and science performance, 4 cybersecurity tests, and 2 safety evaluations. This chapter lists the 10 performance tests. To compare the old and new models under the same conditions, every value uses "max," the highest reasoning setting. All of these are tests OpenAI chose, and all values were measured by OpenAI (third-party indexes are included because OpenAI put them in its charts).
| Test (what it measures) | GPT-6.2 Astra | GPT-6 Astra | GPT-6.2 Sol | GPT-6.1 Sol |
|---|---|---|---|---|
| Agents' Last Exam v1.1 (long professional tasks) | 77.6%$4.47 | 73.1%$8.99 | 73.0%$1.32 | 72.5%$2.02 |
| GDPval v2 (real consulting and finance work) | 50.4%36K output tokens | 44.2%46K output tokens | 46.9%33K output tokens | 41.5%40K output tokens |
| Harvey LAB-AA v1.1 (legal work, rate meeting all criteria) | 15.0%$6.35 | 8.6%$13.67 | 11.1%$1.92 | 6.9%$2.48 |
| FrontierCode 2.0 (real code changes) | 44.2%$5.97 | 42.8%$6.68 | 43.8%$1.29 | 39.0%$1.03 |
| Terminal-Bench-Science 0.1 (scientific work in a terminal) | 77.1%$7.15 | 68.1%$23.80 | 68.0%$2.21 | 57.0%$5.47 |
| Vals Science Index v1 (biology data interpretation and experiments) | 68.4%$4.17 | 65.1%$8.03 | 57.8%$1.03 | 59.6%$1.36 |
| BenchCAD 2.1 Core (CAD design) | 82.5%$5.43 | 62.6%$9.71 | 73.1%$1.60 | 52.6%$1.61 |
| AutomationBench 1.0.6 (business automation) | 52.4%$0.83 | 41.4%$1.73 | 41.1%$0.27 | 36.1%$0.30 |
| LifeVal (difficult everyday tasks) | 86.1%$4.23 | 79.8%$9.04 | 83.8%$1.27 | 79.5%$1.27 |
| Artificial Analysis Intelligence Index v5.0.1 (third-party overall index) | 62.1$4.00 | 56.9$6.24 | 57.3$1.01 | 54.6$1.27 |
Values at each model's max reasoning setting. The small text is the cost per task in US dollars, except for GDPval v2, where it is the output token count per task. Green marks a higher score than the previous model, and red marks a lower one. Source: the chart data in OpenAI's "Introducing GPT-6.2 Astra and Sol" (checked November 19, 2026). GDPval v2 is an OpenAI internal test. The four cybersecurity tests (DeepBinary, ExploitBench, SecureSWE, and PatchBench) are not included. Two of them were measured with safety measures removed in an internal setup, which differs from normal use. The two safety evaluations are covered in Chapter 8.
Three things stand out.
① On all 10 tests, Astra scored higher than the previous Astra, and on many it did so at about half the cost. On Agents' Last Exam, which measures long professional tasks, it scores 77.6%, above GPT-6 Astra's 73.1%, and its cost per task is $4.47, about half. The announcement's headline figure was 78.3% ($2.89) at "xhigh," a setting one step below max. That is 5.2 points higher than the previous Astra at max, and the cost was less than a third. For some tests Astra scores higher at xhigh than at max, so max is not always the best choice.
The largest gains are in legal work (Harvey LAB-AA, 8.6% → 15.0%), CAD design (BenchCAD, 62.6% → 82.5%), and scientific work in a terminal (Terminal-Bench-Science, 68.1% → 77.1%). Harvey LAB-AA counts a result as passing only if it meets every grading criterion and contains no serious fabricated content, so all models score low on it.
② Sol went up on many tests at the same price. On FrontierCode 2.0, which measures whether real code changes would pass a maintainer's merge standards, 6.2 Sol scores 43.8%. That is only 0.4 points behind 6.2 Astra (44.2%), and it costs about one-fifth as much per task. Sol also rose significantly from 6.1 Sol on CAD design and legal work. On the other hand, on the Vals Science Index, which covers biology data, Sol scored 57.8% at max, lower than 6.1 Sol's 59.6% (58.5% at xhigh).
③ Some tests drop sharply at the low reasoning setting. Looking at the low values in the same charts, 6.2 Sol scores 25.3% on LifeVal, an everyday task test (6.1 Sol at low scores 66.8%), and 43.4% on Terminal-Bench-Science. 6.2 Astra at low scores 46.3% on LifeVal and 56.3% on Terminal-Bench-Science. At medium, LifeVal recovers to 70.3% for 6.2 Sol and 78.9% for 6.2 Astra. OpenAI's migration guide also recommends starting at medium when you first use the models. If you set low to reduce costs, first check that your own work does not lose quality.
4. Performance ②: compared with Claude and Gemini
Some of OpenAI's charts also include Anthropic's Claude Fable 5.2 and Claude Opus 5.5, and Google's Gemini 4 Carbon. Because the other companies' models were tested at different reasoning levels, this table lists each model's highest-scoring setting. We left out LifeVal, which does not include the other companies' models. Claude Sonnet 5.5, which has the same price as Sol, does not appear in any of the charts.
| Test | GPT-6.2 Astra | GPT-6.2 Sol | Claude Fable 5.2 | Claude Opus 5.5 | Gemini 4 Carbon |
|---|---|---|---|---|---|
| Agents' Last Exam v1.1 | 78.3%$2.89 · xhigh | 73.0%$1.32 · max | 72.8%$60.05 · max | 71.6%$17.57 · max | 77.4%$4.60 · high |
| GDPval v2 (internal) | 50.4%36K output tokens · max | 46.9%33K output tokens · max | 44.8%222K output tokens · max | 43.5%191K output tokens · max | 48.8%109K output tokens · high |
| Harvey LAB-AA v1.1 | 15.0%$6.35 · max | 11.1%$1.92 · max | 7.2%$39.84 · max | 4.2%$17.93 · max | — |
| FrontierCode 2.0 | 44.2%$5.97 · max | 43.8%$1.29 · max | 49.1%$14.19 · xhigh | 45.5%$12.82 · max | 41.1%$5.16 · medium |
| Terminal-Bench-Science 0.1 | 77.1%$7.15 · max | 68.0%$2.21 · max | 73.3%$45.21 · max | 63.3%$23.21 · max | — |
| Vals Science Index v1 | 68.4%$4.17 · max | 58.5%$0.71 · xhigh | 63.7%$8.79 · high | 58.7%$8.71 · max | — |
| BenchCAD 2.1 Core | 82.5%$5.43 · max | 73.1%$1.60 · max | 83.5%$50.77 · max | 65.3%$20.00 · max | — |
| AutomationBench 1.0.6 | 52.4%$0.83 · max | 41.2%$0.21 · high | 51.5%$4.64 · max | 42.5%$1.44 · max | 53.9%$2.12 · high |
| Artificial Analysis Intelligence Index v5.0.1 | 62.1$4.00 · max | 57.3$1.01 · max | 64.6$26.54 · max | 59.1$10.04 · max | 60.0$7.76 · high |
Values at the reasoning setting where each model scored highest (the setting is written in the small text). "—" means the model is not in the chart. Source is the same as the table in the previous chapter. For Claude Fable 5.2 on Agents' Last Exam, the result excludes 9 tasks that were answered by a different model (a fallback), as noted in the announcement. The announcement does not say which price (before or after the introductory price ends on November 30, 2026) was used to calculate Gemini 4 Carbon's cost. For Claude Fable 5.2 on Terminal-Bench-Science, the value is one Artificial Analysis measured using a setup called mini-swe-agent. Anthropic reports 74.6% with max reasoning in Claude Code (footnote in the announcement).
On these nine tests, 6.2 Astra had the highest score on five. On the other four, Claude Fable 5.2 led on real code changes (FrontierCode 2.0), CAD design (BenchCAD), and the third-party overall index (Artificial Analysis), and Gemini 4 Carbon led on business automation (AutomationBench). In head-to-head comparisons with Fable 5.2, Astra is ahead on six tests and Fable 5.2 on three. The biggest difference is cost. On the eight tests where costs are shown, Fable 5.2's cost per task is about 2 to 21 times Astra's (the 21 times is on Agents' Last Exam, where Astra is at xhigh). Fable 5.2 and Astra have the same price per token, so the main reason is likely the difference in how many tokens each model uses for the same work. However, the reasoning settings compared and the test methods (such as how Terminal-Bench-Science is run) are not identical.
This comparison uses the tests and settings OpenAI chose, and includes only the other companies' models that appear in OpenAI's charts. Anthropic's own announcement uses different tests, and the ranking would look different there. For Claude figures, see Claude Fable 5.2 features and pricing. For Gemini, see Gemini 4 Carbon vs. Argon.
5. Astra or Sol: which to use
Astra's price per token is four times Sol's. Even compared by cost per task, OpenAI's measurements put Astra at about 3 to 5 times Sol's cost (both at max). The deciding question is whether the gain in output quality is worth the difference in cost for your work.
OpenAI's API list says "If you're unsure, use Astra," but this article recommends "If you're unsure, use Sol." The reason is that the cost difference is 3 to 5 times, while the gap is small on real code changes (FrontierCode 2.0) and everyday tasks (LifeVal). Here is a guide by use case.
- Everyday coding, writing, and document work → Sol On the real code change test, Astra is only 0.4 points ahead, and it costs about one-fifth as much. New chats in ChatGPT Work also start with Sol. Start with Sol when unsure, and switch to Astra only for the cases where it falls short. This is the least wasteful approach.
- Business automation, long and complex professional work, finishing slides and spreadsheets → Astra On business automation (AutomationBench), Astra scores 52.4% and Sol 41.1%, a gap of about 11 points. Professional long tasks (Agents' Last Exam), legal work (Harvey LAB-AA), and GDPval v2 show gaps of 3.5 to 4.6 points. OpenAI says it trained Astra to produce slides, spreadsheets, and documents that follow templates and the conventions of each field.
- Scientific calculation and CAD design → Astra Astra scores 77.1% on Terminal-Bench-Science against Sol's 68.0%, and 82.5% on BenchCAD against 73.1%. These are areas where the gap is large.
- Fast and cheap high-volume work without reasoning → GPT-6.1 Terra or GPT-6 Luna Both GPT-6.2 models cannot set reasoning to "none."
- If you use Claude Fable 5.2 has the same price as Astra, and Sonnet 5.5 has the same price as Sol. In OpenAI's measurements, Fable 5.2 led on real code changes and CAD design, while Astra led on many other work tasks, and Astra's cost came out much lower. The most reliable approach is to run your usual work through both a few times each and compare the revisions you need to make and the bill.
For how the new models differ from the previous Astra and Sol, see the GPT-6 Astra explainer and the comparison of GPT-6.1 Sol and GPT-6 Sol.
6. Using them in ChatGPT, Codex, and GitHub Copilot
ChatGPT and Codex
- Eligible plans: The announcement covers all Plus, Pro, Enterprise, and Edu users. Free is not included.
- Where to use them: ChatGPT Work and Codex. ChatGPT's release notes (November 17) describe the models as a "preview" in the Work and Codex desktop apps. You choose them from the model selector in Work or in Codex. The options in regular Chat (conversations outside Work) do not change.
- Default model: According to ChatGPT's documentation, new chats in Work start with 6.2 Sol (reasoning strength Light for individual plans, and Medium for Enterprise). Sol uses fewer credits per token than Astra.
- Ultrafast: Pro 500 can use Ultrafast with both models. Ultrafast becomes available on Pro 200 for the first time, starting with 6.2 Sol (the announcement does not say when Astra will be added). The announcement says nothing about Pro 100. Eligible Enterprise and Edu workspaces can use the same feature once a workspace admin enables it.
- Codex CLI: ChatGPT's Models documentation gives
codex --model gpt-6.2-solas an example of a command for specifying a model in Codex.
Sources: the "Pricing and availability" section of OpenAI's announcement, ChatGPT release notes (entry for November 17, 2026), and the ChatGPT Models documentation (checked November 19, 2026). In managed workspaces, the models you see depend on your seat type and settings.
For usage limits across the three Pro plans, see ChatGPT Pro 100, 200, and 500 compared. For Work itself, see What is ChatGPT Work?
GitHub Copilot
On November 17, GitHub made the two models generally available in Copilot. They are available to Copilot Pro+, Max, Business, and Enterprise, and you can select them in the model picker in VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI, and github.com. Pricing is usage-based and calculated at OpenAI's list prices. For Business and Enterprise, the new models are enabled automatically unless an administrator has disabled them. The rollout is gradual, so some users may not see the models in their list yet (GitHub Changelog).
7. What to check when moving to the API
Changing the model name to gpt-6.2-astra or gpt-6.2-sol is enough to get things running. However, OpenAI's migration guide, Using GPT-6.2, lists the following points to check when you switch. The first two are changes from GPT-6.2 that affect quality and cost if you miss them.
- Rewriting part of a conversation discards the earlier reasoning With the Responses API, if you rewrite earlier parts of the conversation history (such as earlier developer messages or tool calls), GPT-6.2 no longer uses the stored reasoning records (reasoning items) from before that point. The same applies when you switch to a smaller model, or when the end user or organization changes. Losing reasoning can lower answer quality, and redoing earlier work can increase the number of tokens used. When you change settings, the basic approach is to leave the earlier parts unchanged and add the new instructions at the end.
- Put shared instructions at the start of the input, not in
instructionsFrom GPT-6.2, the outermostinstructionsfield in the Responses API is placed afterinput. If you want the same instructions to be cached on every request, put them in the first item ofinputas a developer message. Useinstructionsonly for instructions specific to that one request.
Example:{"model": "gpt-6.2-sol", "input": [{"role": "developer", "content": "Answer according to the internal glossary"}, {"role": "user", "content": "…"}]} - Start at medium reasoning When you start using the models, OpenAI recommends trying medium first. If you currently use
noneorminimal, switch tolow(none is not accepted). However, as Chapter 3 showed, some tests drop sharply at low, so confirm quality on your own work before deciding. - Use the Responses API if you use tools You can use Chat Completions, but tool calls require the Responses API.
- Remove unsupported parameters Remove
temperature,top_p, andtop_logprobs(andlogprobsin Chat Completions). - Cache settings You do not need to change anything when moving from the GPT-6 series. Only when moving from GPT-5.5 or earlier, replace
prompt_cache_retentionwithprompt_cache_options.ttl("30m"). For differences in how caching works, see OpenAI and Claude prompt caching compared. - Review your instruction text The migration guide says GPT-6.2 Astra writes shorter and more direct output than before even without instructions to do so (the same statement does not appear for Sol). If you have instructions that strongly ask for brevity, check that the output is not becoming too short. The model may also propose next steps and wait for confirmation when the intent is unclear, so for long tasks, state that it may proceed on reasonable assumptions.
Sources: OpenAI's "Using GPT-6.2" and the API changelog (checked November 19, 2026). The announcement also recommends leaving the earlier parts of a conversation unchanged and adding new instructions at the end.
8. Safety
OpenAI announced that both models reached the "Critical" threshold for cybersecurity and the "High" threshold for biology and chemistry under its own Preparedness Framework. This is the same rating as GPT-6 Astra. OpenAI says it released the models after strengthening restrictions on high-risk cyber assistance and certain biology requests. As a result, some legitimate vulnerability research and security testing may also be blocked. OpenAI says it will work toward bringing GPT-6.2 to Daybreak, its cyber defense program for defenders, but has not announced a date.
Also, 6.2 Astra is the first model OpenAI has released publicly that reached "High" for AI self-improvement, meaning its ability to speed up AI research. OpenAI says it has set up measures such as monitoring research work and stopping outputs when necessary. In the announcement's chart of inappropriate behavior in workplace tasks, the rate of inappropriate actions in tasks modeled on a workplace was 0.64% for 6.2 Astra and 0.71% for 6.2 Sol, both lower than GPT-6 Astra's 1.2%. On the other hand, two tests in the announcement's chart on concealing failures showed a stronger tendency to not report obstacles such as missing files or unavailable search tools. OpenAI says it reduced this by adding instructions to Codex. For details, see OpenAI's GPT-6.2 System Card.
Summary
- GPT-6.2 Astra and Sol were released on November 17, 2026 (US time) as successors to GPT-6 Astra and GPT-6.1 Sol. Astra is the top model, and Sol is the main model for everyday work.
- API prices: Astra is $8 input and $40 output (20% cheaper than GPT-6 Astra, cache 60% cheaper). Sol is $2 input and $10 output (unchanged).
- In OpenAI's measurements, both score higher than their predecessors on many tests, and Astra's cost per task fell to about half on many tests.
- Claude Fable 5.2 leads Astra on real code changes, CAD design, and the overall index, but its cost per task is about 2 to 21 times Astra's (OpenAI's measurements).
- When unsure, start with Sol, and try medium reasoning first (some tests drop sharply at low).
- In the API, check two things: that rewriting earlier parts of a conversation discards reasoning, and that the position of
instructionshas changed.
FAQ
Q. How much do the GPT-6.2 Astra and Sol API prices cost?
A. Per 1 million tokens, Astra costs $8 input, $0.40 cached input, and $40 output. Sol costs $2 input, $0.10 cached input, and $10 output. Both models have Batch at half price and Fast at 2 times. Ultrafast is 3 times until February 28, 2027, and 6 times from March 1. For input over 272K tokens, the input and cache for the entire request are 2 times the normal price, and output is 1.5 times.
Q. What is the difference between GPT-6.2 Sol and GPT-6.1 Sol?
A. The price is the same. In OpenAI's measurements, 6.2 Sol scores higher on many tests (real code changes went from 39.0% to 43.8%, and CAD design from 52.6% to 73.1%). Its knowledge cutoff moved from April 30, 2026 for 6.1 Sol to June 1, 2026 for 6.2 Sol. On the Vals Science Index, which covers biology data, 6.1 Sol scored higher at max.
Q. Is it better than Claude Fable 5.2?
A. It depends on the test. On the nine tests where OpenAI shows both, 6.2 Astra scored higher on six, and Fable 5.2 scored higher on three: real code changes, CAD design, and the third-party overall index. On the eight tests where costs are shown, Fable 5.2's cost per task was about 2 to 21 times Astra's (comparing each model at the setting where it scored highest). This is a comparison measured by OpenAI.
Q. Can I use it on the free ChatGPT plan?
A. The announcement covers Plus, Pro, Enterprise, and Edu. Free is not included. You can use the models in ChatGPT Work and Codex, and the options in regular Chat do not change.
Q. Should I switch from GPT-6 Astra?
A. In many cases, yes. The price is 20% lower, and in OpenAI's measurements it beats GPT-6 Astra on all 10 performance tests. However, the tests were chosen by OpenAI, and there are changes to consider, such as shorter output and the inability to set reasoning to "none." In the API, rewriting earlier parts of a conversation discards reasoning, and the handling of instructions has changed. Check the items in Chapter 7 before switching.
Update log
- November 19, 2026First version. Created by checking the chart data from OpenAI's announcement (November 17), the API model pages and migration guide, ChatGPT's release notes and documentation, and the GitHub Changelog.
Sources
- OpenAI: Introducing GPT-6.2 Astra and Sol, GPT-6.2 Astra (API), GPT-6.2 Sol (API), Models, Using GPT-6.2, API changelog, ChatGPT release notes, ChatGPT Models
- GitHub: OpenAI's GPT-6.2 Astra and GPT-6.2 Sol now available
- OpenAI: GPT-6.2 System Card
- Comparison prices (checked November 19, 2026): Anthropic Pricing, Google Gemini API pricing