LLM prices, quality, speed and hallucination rates compared
Prices come from each vendor's own pricing page. Quality is the Artificial Analysis Intelligence Index, a composite of hard tasks with checkable answers. Speed is output tokens per second as Artificial Analysis measured it. Hallucination is the Vectara leaderboard rate (updated 22 September 2026), the share of summaries that add claims not in the source; lower is better. An asterisk marks a rate from an earlier version of that model. Gemma has no paid API price.
| Model | Vendor | Quality index | Output tokens/s | Hallucination rate | Input $/M | Output $/M | Cached read $/M |
|---|
| Haiku 5.5 | Anthropic | 38 | 175 | 9.8%* | $0.1 | $0.5 | $0.01 |
| Sonnet 5.5 | Anthropic | 47 | 96 | 10.6%* | $2 | $10 | $0.1 |
| Opus 5.5 | Anthropic | 54 | 82 | 12%* | $4 | $20 | $0.2 |
| Fable 5.1 | Anthropic | 51 | 52 | no data | $10 | $50 | $0.25 |
| GPT-6 Astra | OpenAI | 51 | 55 | 8.7% | $10 | $50 | $1 |
| GPT-6.1 Sol | OpenAI | 50 | 57 | 6.5%* | $2 | $10 | $0.1 |
| GPT-5.6 Terra | OpenAI | 34 | 102 | no data | $2 | $12 | $0.2 |
| GPT-6 Luna | OpenAI | 33 | 134 | no data | $0.1 | $0.5 | $0.01 |
| GLM-5.3 | Z.ai | 45 | 66 | 10.1%* | $1.4 | $4.4 | $0.26 |
| GLM-5.3-Flash | Z.ai | 42 | 62 | 9.3%* | $0.15 | $0.5 | $0.03 |
| Qwen3.8-Max | Alibaba | 45 | 41 | no data | $2 | $6 | none listed |
| Qwen3.7-Plus | Alibaba | 25 | 55 | 10.7%* | $0.4 | $1.6 | none listed |
| Qwen3.8 27B | Alibaba | 34 | 46 | 12.1%* | $0.5 | $3 | none listed |
| Gemini 3.8 Flash | Google | 41 | 137 | 13.5%* | $0.75 | $3.75 | $0.075 |
| Gemini 3.5 Flash-Lite | Google | 22 | 408 | 8.2%* | $0.3 | $2.5 | $0.03 |
| Gemma 4 31B | Google (open) | 15 | 35 | 7.4% | free tier | free tier | - |
| Gemma 4 26B A4B | Google (open) | 17 | 75 | 5.2% | free tier | free tier | - |
Sources: Anthropic pricing, OpenAI pricing, Z.ai pricing, Alibaba pricing, Google pricing, Artificial Analysis leaderboard, Vectara hallucination leaderboard.
Read the questions and answers
Last verified 14 October 2026. Prices and rankings change, so check each vendor's page before you rely on a figure. By Ali Ameen.
Which AI model should you use for your language, framework, app or website?
How these answers are made. No public benchmark measures a model on one language, framework or kind of site, so each answer applies a stated rule to the data in the table above. A task is rated light (minimum quality score 33), standard (47) or hard (50) on the Artificial Analysis Intelligence Index. The value pick is the cheapest model, by a blend of three parts input price to one part output price, within 15% of the lowest price among models at or above that minimum. The strongest is the highest score. Speed and hallucination figures are added where a task calls for them. The language-specific advice is general engineering practice, not benchmark results. Verified 14 October 2026.
Best AI model by programming language
Which is the best AI model for Python?
For Python, GPT-6.1 Sol or Sonnet 5.5 is the value pick, and Opus 5.5 is the strongest. No public benchmark scores Python on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. A score of 47 or higher covers everyday work. Escalate to Opus 5.5 (54) for hard cases. A cheaper option just below that line is GLM-5.3-Flash (quality 42, $0.15 input and $0.5 output per million tokens). Python is the most heavily tested language in public coding benchmarks (SWE-bench Verified uses Python repositories), so general scores carry over to it better than to other languages. Run 30 of your own Python tasks through two models before you commit.
Which is the best AI model for TypeScript?
For TypeScript, GPT-6.1 Sol or Sonnet 5.5 is the value pick, and Opus 5.5 is the strongest. No public benchmark scores TypeScript on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. A score of 47 or higher covers everyday work. Escalate to Opus 5.5 (54) for hard cases. A cheaper option just below that line is GLM-5.3-Flash (quality 42, $0.15 input and $0.5 output per million tokens). Ask for strict mode and run tsc on every answer. Type errors are the quickest way to spot a weak answer. Run 30 of your own TypeScript tasks through two models before you commit.
Which is the best AI model for JavaScript?
For JavaScript, GPT-6.1 Sol or Sonnet 5.5 is the value pick, and Opus 5.5 is the strongest. No public benchmark scores JavaScript on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. A score of 47 or higher covers everyday work. Escalate to Opus 5.5 (54) for hard cases. A cheaper option just below that line is GLM-5.3-Flash (quality 42, $0.15 input and $0.5 output per million tokens). State the runtime (browser or Node.js) and the module system (ESM or CommonJS), because models mix them. Run 30 of your own JavaScript tasks through two models before you commit.
Which is the best AI model for Java?
For Java, GPT-6.1 Sol or Sonnet 5.5 is the value pick, and Opus 5.5 is the strongest. No public benchmark scores Java on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. A score of 47 or higher covers everyday work. Escalate to Opus 5.5 (54) for hard cases. A cheaper option just below that line is GLM-5.3-Flash (quality 42, $0.15 input and $0.5 output per million tokens). Name the Java version and the build tool (Maven or Gradle). Answers often default to older Java. Run 30 of your own Java tasks through two models before you commit.
Which is the best AI model for C#?
For C#, GPT-6.1 Sol or Sonnet 5.5 is the value pick, and Opus 5.5 is the strongest. No public benchmark scores C# on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. A score of 47 or higher covers everyday work. Escalate to Opus 5.5 (54) for hard cases. A cheaper option just below that line is GLM-5.3-Flash (quality 42, $0.15 input and $0.5 output per million tokens). State the .NET version, for example .NET 8 or later versus .NET Framework, because the APIs differ. Run 30 of your own C# tasks through two models before you commit.
Which is the best AI model for C++?
For C++, GPT-6.1 Sol is the value pick, and Opus 5.5 is the strongest. No public benchmark scores C++ on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. This work rewards deeper reasoning, so start at a score of 50 or higher and escalate to Opus 5.5 (54) when an answer fails to compile or pass tests. A cheaper option just below that line is GLM-5.3 (quality 45, $1.4 input and $4.4 output per million tokens). Wrong answers are costly here. On the Vectara leaderboard, GPT-6.1 Sol has the lowest hallucination rate among the qualifying models with data, 6.5% (measured on GPT-6 Sol, earlier version), though that test covers summaries of supplied documents, not code. Compile with warnings as errors and run the sanitizers (ASan and UBSan) on generated code, because undefined behaviour still compiles. Run 30 of your own C++ tasks through two models before you commit.
Which is the best AI model for Go?
For Go, GPT-6.1 Sol or Sonnet 5.5 is the value pick, and Opus 5.5 is the strongest. No public benchmark scores Go on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. A score of 47 or higher covers everyday work. Escalate to Opus 5.5 (54) for hard cases. A cheaper option just below that line is GLM-5.3-Flash (quality 42, $0.15 input and $0.5 output per million tokens). Run go vet and the race detector (go test -race) on generated concurrent code. Run 30 of your own Go tasks through two models before you commit.
Which is the best AI model for Rust?
For Rust, GPT-6.1 Sol is the value pick, and Opus 5.5 is the strongest. No public benchmark scores Rust on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. This work rewards deeper reasoning, so start at a score of 50 or higher and escalate to Opus 5.5 (54) when an answer fails to compile or pass tests. A cheaper option just below that line is GLM-5.3 (quality 45, $1.4 input and $4.4 output per million tokens). Paste the full compiler error back into the model rather than describing it. The borrow checker's message is the best prompt. Run 30 of your own Rust tasks through two models before you commit.
Best AI model by framework
Which is the best AI model for FastAPI?
For FastAPI, GPT-6.1 Sol or Sonnet 5.5 is the value pick, and Opus 5.5 is the strongest. No public benchmark scores FastAPI on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. A score of 47 or higher covers everyday work. Escalate to Opus 5.5 (54) for hard cases. A cheaper option just below that line is GLM-5.3-Flash (quality 42, $0.15 input and $0.5 output per million tokens). Ask for Pydantic v2 models and typed dependencies, and state the versions, because Pydantic changed between v1 and v2. Run 30 of your own FastAPI tasks through two models before you commit.
Which is the best AI model for React?
For React, GPT-6.1 Sol or Sonnet 5.5 is the value pick, and Opus 5.5 is the strongest. No public benchmark scores React on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. A score of 47 or higher covers everyday work. Escalate to Opus 5.5 (54) for hard cases. A cheaper option just below that line is GLM-5.3-Flash (quality 42, $0.15 input and $0.5 output per million tokens). State the React version and paste the component tree for context. For visual quality, see the Arena WebDev leaderboard. Run 30 of your own React tasks through two models before you commit.
Which is the best AI model for Next.js?
For Next.js, GPT-6.1 Sol or Sonnet 5.5 is the value pick, and Opus 5.5 is the strongest. No public benchmark scores Next.js on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. A score of 47 or higher covers everyday work. Escalate to Opus 5.5 (54) for hard cases. A cheaper option just below that line is GLM-5.3-Flash (quality 42, $0.15 input and $0.5 output per million tokens). State whether you use the App Router or the Pages Router, and the Next.js version. The two differ substantially. Run 30 of your own Next.js tasks through two models before you commit.
Which is the best AI model for Three.js?
For Three.js, GPT-6.1 Sol or Sonnet 5.5 is the value pick, and Opus 5.5 is the strongest. No public benchmark scores Three.js on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. A score of 47 or higher covers everyday work. Escalate to Opus 5.5 (54) for hard cases. A cheaper option just below that line is GLM-5.3-Flash (quality 42, $0.15 input and $0.5 output per million tokens). State the three.js release, paste the documentation for any API it uses, and test in the browser. Class names and import paths change across releases. Run 30 of your own Three.js tasks through two models before you commit.
Best AI model by type of app or site
Which is the best AI model for a mobile app?
For a mobile app, GPT-6.1 Sol or Sonnet 5.5 is the value pick, and Opus 5.5 is the strongest. No public benchmark scores a mobile app on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. A score of 47 or higher covers everyday work. Escalate to Opus 5.5 (54) for hard cases. A cheaper option just below that line is GLM-5.3-Flash (quality 42, $0.15 input and $0.5 output per million tokens). Pick one platform and framework first, and state the minimum OS versions. Run 30 of your own a mobile app tasks through two models before you commit.
Which is the best AI model for an admin dashboard?
For an admin dashboard, GPT-6.1 Sol or Sonnet 5.5 is the value pick, and Opus 5.5 is the strongest. No public benchmark scores an admin dashboard on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. A score of 47 or higher covers everyday work. Escalate to Opus 5.5 (54) for hard cases. A cheaper option just below that line is GLM-5.3-Flash (quality 42, $0.15 input and $0.5 output per million tokens). Give the data shape and the permissions model first. Charts are easy; access rules are not. Run 30 of your own an admin dashboard tasks through two models before you commit.
Which is the best AI model for an online store?
For an online store, GPT-6.1 Sol or Sonnet 5.5 is the value pick, and Opus 5.5 is the strongest. No public benchmark scores an online store on its own, so this answer uses the Artificial Analysis Intelligence Index as the nearest general measure: the models compared here score from 15 to 54, and GPT-6.1 Sol scores 50 at $2 input and $10 output per million tokens. A score of 47 or higher covers everyday work. Escalate to Opus 5.5 (54) for hard cases. A cheaper option just below that line is GLM-5.3-Flash (quality 42, $0.15 input and $0.5 output per million tokens). Wrong answers are costly here. On the Vectara leaderboard, GPT-6.1 Sol has the lowest hallucination rate among the qualifying models with data, 6.5% (measured on GPT-6 Sol, earlier version), though that test covers summaries of supplied documents, not code. Keep payment code on the provider's documented SDK and review every checkout change by hand. Run 30 of your own an online store tasks through two models before you commit.