Skip to main content
<- Articles

10 Best Reasoning Models in 2026 and What They Actually Cost

Gemini 4 Neon tops the reasoning ranking at under half of Claude Fable 5.2's price, with Fable and GPT-6.2 Astra close behind. Compare the top ten on reasoning score, effort settings, and real cost.

Respan · October 22, 2026 · Updated November 20, 2026 · 29 min read

Respan Gateway: up to 55% off tokens on Gemini, GPT, Claude and 1,000+ more models. One API key, 0% fee.

Thinking is now the default at the top of the model market. Claude Fable 5.2, GPT-6.2 Astra, Gemini 4 Carbon, Claude Opus 5.5, and Kimi K3.1 Pro all reason on every request, and none of them offers a setting that turns reasoning off.

That thinking has a price. Reasoning tokens bill like output, often without showing up in the response, and default effort settings range from medium on some models to max on others. Because of that, two models with similar per-token rates can produce very different bills for the same request.

The right pick is the cheapest model, at the lowest effort level, that still clears your quality bar on your own tasks. Getting there means looking past headline scores to how each model spends and bills its thinking.

These are the 10 best AI reasoning models in 2026 and what they actually cost to run.

TL;DR: Best AI reasoning models at a glance

Ranked by reasoning score, each model fits a different kind of work and a different reasoning budget:

  • Gemini 4 Neon - The top reasoning score (93.1), at $3/$15 per 1M tokens with thinking set to high by default.
  • Claude Fable 5.2 - Second at 90.0, for the hardest long-horizon work, at $8/$40 per 1M tokens with effort set to high by default.
  • GPT-6.2 Astra - OpenAI's top tier at 89.8, at the same $8/$40 as Fable 5.2, with no way to turn reasoning off.
  • Gemini 4 Carbon - Long, multi-step agent runs, spending more thinking tokens than Argon at the same price.
  • Claude Opus 5.5 - Anthropic's recommended default for coding and knowledge work, with effort set to medium by default.
  • GPT-6.1 Terra - Low-cost reasoning at $1/$5 per 1M tokens, with reasoning you can turn off.
  • Kimi K3.1 Pro - The only open-weight model in the top 10, running at max effort by default.
  • Muse Wonder Full - The larger Muse Wonder model, available only in Muse Code, with Muse Wonder as the version you can call.
  • Claude Sonnet 5.5 - Coding-heavy work, with thinking you can limit to the steps between tool calls.
  • Gemini 4 Argon - Long-horizon work on fewer thinking tokens than Carbon at the same price, with medium thinking by default.

For agents where most calls are simple, a model you can dial down (Terra with reasoning off, or Sonnet 5.5 thinking only between tool calls) can handle both the simple steps and the hard ones from a single model ID. Respan's AI gateway routes every call through one API to 1,000+ models and prices each request, so you can see what each step costs and route it to the model it needs.

What are AI reasoning models?

AI reasoning models are large language models that work through a problem in intermediate steps before writing their final answer. Instead of producing a response in one pass, they generate a stretch of reasoning first, which lets them break a task down, test an approach, and correct course before committing to an output.

The line between a reasoning model and a flagship model has mostly closed, since the top models from Anthropic, OpenAI, Google, xAI, and Moonshot all think by default, so among the types of AI models developers choose between, reasoning is more often a setting than a separate model. Overall capability is the lens for ranking the best frontier AI models. For reasoning models, the deciding factors are how much each one thinks and what that thinking costs you.

How do AI reasoning models work?

Reasoning models spend extra compute at inference time, and the provider gives you a dial for how much.

Thinking before answering

A reasoning model generates a chain of intermediate tokens before it writes the answer. In that chain it breaks the problem into steps, tries an approach, checks it, and revises. Labs train this behavior with reinforcement learning on tasks with verifiable outcomes, such as math problems with known answers and code that has to pass tests, so the model learns that spending more tokens on a hard problem raises the odds of getting it right.

More thinking usually means better answers on multi-step problems. It also means higher latency and a bigger bill, which is why providers give you some control over how much their models think.

Effort levels and thinking budgets

Every provider here exposes an effort parameter with named levels, but the level names and defaults don't line up across APIs. If you never set the parameter, Kimi K3.1 Pro and Muse Wonder run at max effort and Claude Fable 5.2, Sonnet 5.5, and Gemini 4 Neon at high, while Claude Opus 5.5, GPT-6.1 Terra, Gemini 4 Carbon, and Gemini 4 Argon run at medium. OpenAI doesn't state a default for GPT-6.2 Astra.

ModelParameterValuesDefaultOff switchDocs
Gemini 4 Neonthinking_levelFour levels, up to xhighhighNot statedNeon announcement
Claude Fable 5.2output_config.effortlow, medium, high, xhigh, maxhighNoAnthropic's effort docs
GPT-6.2 Astrareasoning.effortNot listedNot statedNo (none rejected)GPT-6.2 Astra model page
Gemini 4 Carbonthinking_levellow, medium, highmediumNoCarbon developer guide
Claude Opus 5.5output_config.effortlow, medium, high, xhigh, maxmediumNoSame docs as Fable 5.2
GPT-6.1 Terrareasoning.effortnone, low, medium, high, xhigh, maxmediumnoneGPT-6.1 Terra model page
Kimi K3.1 Proreasoning_effortlow, high, maxmaxNoKimi reasoning effort
Muse Wonderreasoning.effortminimal, low, medium, high, xhigh, maxmaxNoMeta's reasoning docs
Claude Sonnet 5.5output_config.effortlow, medium, high, xhigh, maxhighbetween_tools (partial)Same docs as Fable 5.2
Gemini 4 Argonthinking_levellow, medium, highmediumNoGemini thinking docs

Set the parameter explicitly in production so a provider's default change doesn't move your costs. Anthropic lowered the default from high on Opus 5 to medium on Opus 5.5, and a request that left effort unset changed behavior with the model upgrade.

How reasoning tokens are billed

Anthropic, OpenAI, Google, and Meta state that thinking tokens bill at the output rate, and Moonshot counts reasoning toward token usage as well.

What you see varies more than what you pay. OpenAI and Meta keep raw reasoning private and return summaries only when you opt in, Gemini does the same with thought summaries, and Claude Opus 5.5 omits thinking text by default. Kimi K3.1 Pro returns its reasoning in a reasoning_content field. Anthropic notes that thinking bills in full even when the text isn't returned, and its summarized display still bills the complete thinking, so the token count on your invoice won't match the text in your logs.

What are reasoning models typically used for?

Reasoning models earn their cost on tasks where the answer depends on intermediate steps the model has to get right:

  • Multi-step agents and tool calling - An agent that plans, calls a tool, reads the result, and adjusts gets better results when it can reason between calls. Claude Opus 5.5 interleaves thinking between tool calls automatically, and Google built Gemini 4 Carbon for long build-and-verify loops.
  • Coding and debugging - Tracing a bug across files, planning a refactor, or fixing a failing test suite. Most labs now lead their launch numbers with agentic coding benchmarks such as DeepSWE and Terminal-Bench.
  • Math and science problems - Multi-step derivations and graduate-level science questions, the territory benchmarks like GPQA Diamond and Humanity's Last Exam measure.
  • Long-document analysis - Contracts, codebases, and research corpora that fill a 1M-token context window, where the model has to connect facts that sit far apart.

In each of these, a wrong intermediate step produces a wrong final answer, which is the failure that extra thinking is built to catch.

When not to use a reasoning model

A lot of the calls inside a production app don't need the model to think, and paying for reasoning on them adds cost and latency without improving the output:

  • Extraction and classification - Pulling fields from a document or labeling a support ticket has one right answer that doesn't require planning.
  • Latency-sensitive responses - Voice agents, autocomplete, and live chat feel slow when thinking delays the first visible token.
  • Simple steps inside an agent - Formatting a tool result, summarizing a short page, or picking the next tool from a short list.
  • High-volume batch jobs with fixed formats - Rewriting product descriptions or tagging records at scale, where small per-call savings add up across millions of requests.

On models that let you dial reasoning down, the fix is a parameter change: none on GPT-6.1 Terra, or thinking: {"type": "between_tools"} on Claude Sonnet 5.5, which keeps thinking only between tool calls. For models that always think, an LLM router can send those calls to a cheaper non-reasoning model instead.

10 best reasoning models in 2026

Five labs place models in the top 10 on reasoning, and below the top three the scores sit close enough together that effort defaults and pricing separate them as much as score does.

How we ranked them

Every model below is ranked by its Respan Index reasoning score, which blends public benchmarks and lab-reported results into one 0 to 100 score. Respan's reasoning ranking keeps that list current for models with published reasoning results, while this list also includes models whose reasoning score is estimated, marked with an asterisk. It takes the current model from each line and leaves out versions their labs have replaced, such as Claude Fable 5.1, GPT-6 Astra, and GPT-6.1 Sol.

The benchmark figures inside each entry are the lab's own reported numbers. The labs don't publish the same benchmarks, so a lab-reported score is only comparable to another lab's score on the same test. Prices are per 1M input and output tokens at standard rates.

ModelReasoningOverallDefault effortOff switchPrice (in / out)
Gemini 4 Neon93.180.6HighNot stated$3 / $15
Claude Fable 5.290.080.9HighNo$8 / $40
GPT-6.2 Astra89.875.3Not statedNo$8 / $40
Gemini 4 Carbon84.076.2MediumNo$2 / $10, then $4 / $20
Claude Opus 5.583.976.0MediumNo$4 / $20
GPT-6.1 Terra82.869.1MediumYes$1 / $5
Kimi K3.1 Pro82.8*71.5MaxNo$2 / $10
Muse Wonder Full82.5*71.0Max (Muse Wonder)NoNo API ($4 / $20 for Muse Wonder)
Claude Sonnet 5.578.973.4HighPartial$2 / $10
Gemini 4 Argon78.569.3MediumNo$2 / $10, then $4 / $20

*Estimated score.

1. Gemini 4 Neon

Google released Gemini 4 Neon on November 18 as the Gemini 4 model between Carbon and Hydrogen, and it leads the reasoning ranking at 93.1, 3.1 points ahead of Claude Fable 5.2 at well under half of Fable's price. Its overall score of 80.6 sits 0.3 points behind Fable 5.2's, and Google reports 71.3% on Humanity's Last Exam (HLE-Diamond) and 84.0% on RiemannBench for math.

Neon has four thinking levels, with high as the default and xhigh as the highest, though Google's announcement doesn't name the other two and Neon isn't yet listed in Google's thinking docs, so confirm the level names before you set one. Google says Neon costs 32% less per task than Carbon at its high level, and it's on the paid API tier only. Real-time safeguards target cyber and biosecurity misuse, and a separate Neon for cyber is limited to defenders in Google's Fairwind Program.

  • Reasoning score: 93.1 (#1) · Overall 80.6
  • Reasoning controls: thinking_level, four levels up to xhigh, high by default
  • Pricing: $3 input / $15 output per 1M tokens, with cached input at $0.15
  • Context window: 1M tokens
  • Best for: Hard reasoning steps below Fable 5.2's price

2. Claude Fable 5.2

Anthropic released Claude Fable 5.2 on November 9, and it ranks second on reasoning at 90.0, 3.1 points behind Gemini 4 Neon and 0.2 ahead of GPT-6.2 Astra, with a knowledge score of 94.0 that no other model here comes close to. Anthropic reports 74.2% on Humanity's Last Exam with tools, against 67.7% for Opus 5.5, yet it still recommends starting most coding and knowledge work on Opus 5.5, which is the tradeoff to weigh when you compare Fable and Opus.

Adaptive thinking is always on, and output_config.effort defaults to high on the API, one level above Opus 5.5's medium. Fable 5.2 is the same model as Claude Mythos 5.2 with cyber and biology safeguards added, so classifiers block high-risk cyber work and most research biology tasks and reroute them to Opus 4.8 and Opus 5, though Anthropic reports about 55% fewer blocks than on Fable 5.1. Data is retained for 30 days for safety monitoring unless your account qualifies for zero retention, and at high effort and above Anthropic recommends a large max_tokens, since that limit covers thinking and the response together.

  • Reasoning score: 90.0 (#2) · Overall 80.9
  • Reasoning controls: output_config.effort (low to max), high by default, can't be turned off
  • Pricing: $8 input / $40 output per 1M tokens, $16 / $80 in fast mode, with thinking billed as output
  • Context window: 1M tokens, 128K max output
  • Best for: The hardest long-horizon reasoning, as an escalation tier

3. GPT-6.2 Astra

OpenAI released GPT-6.2 Astra on November 17 as its most capable model for the hardest projects, replacing GPT-6 Astra as one of its four current GPT models, and it ranks third on reasoning at 89.8, 0.2 points behind Fable 5.2 at the same $8/$40 per 1M tokens. OpenAI reports it 9 points above GPT-6 Astra on Terminal-Bench-Science 0.1 at under one-third of the cost, and 5.2 points higher on Agents' Last Exam.

The API rejects reasoning.effort set to none, and OpenAI's model page doesn't list the supported levels or a default, so set one explicitly. Starting with GPT-6.2, the API also drops reasoning items when it detects mid-conversation edits to the history, a model downgrade, or a change of end user or organization, which matters for agents that rewrite earlier turns. Requests over 272K input tokens bill the whole request at 2x input and 1.5x output, and fast mode costs 2x the standard rate. How much of Astra's premium over GPT-6.2 Sol pays off by task is what GPT-6 Sol vs Terra vs Luna vs Astra works through.

  • Reasoning score: 89.8 (#3) · Overall 75.3
  • Reasoning controls: reasoning.effort, levels and default not listed, none rejected
  • Pricing: $8 input / $40 output per 1M tokens, with reasoning billed as output, $0.40 cached input, and a surcharge above 272K input tokens
  • Context window: 1.05M tokens, 128K max output
  • Best for: High-stakes reasoning where a failure costs more than the tokens

4. Gemini 4 Carbon

Google built Gemini 4 Carbon on Gemini 4 Argon to spend more thinking tokens per task, and that shows in the ranking: Carbon scores 84.0 on reasoning against Argon's 78.5, at the same per-token price. Google opened it to paying developers on November 10 and reported a score of 59 on the Artificial Analysis Intelligence Index, first on that index at launch.

Carbon uses the same thinking_level setting as Argon (low, medium, or high), with medium as the default and no off setting, and a minimal level returns an error. Thinking and the answer share one output cap, so a tight max_output_tokens can end a response before any answer appears while the thinking still bills. Its introductory $2/$10 rate runs through November 30 and then doubles, and the Gemini model comparison breaks down where it sits next to Neon, Argon, and the Flash models.

  • Reasoning score: 84.0 (#4) · Overall 76.2
  • Reasoning controls: thinking_level (low, medium, high), medium by default, no off setting
  • Pricing: $2 input / $10 output per 1M tokens through November 30, 2026, then $4 / $20, with thinking billed as output
  • Context window: 1M tokens shared by input and output, 65,536 output by default
  • Best for: Long, multi-step agent runs

5. Claude Opus 5.5

Claude Opus 5.5 can't run without thinking, and it places fifth on reasoning at 83.9, a tenth of a point behind Gemini 4 Carbon. Anthropic's adaptive thinking is always on, and output_config.effort controls how deep it goes rather than whether it happens. Anthropic reports 67.7% on Humanity's Last Exam with tools, 66.4% on Terminal-Bench 4.0 at xhigh effort, and 81.8% on OSWorld 2.1. Interleaved thinking between tool calls needs no extra setup, which suits agents that plan across long chains of tool calls.

Forced tool use isn't supported (tool_choice accepts only auto or none), and changing temperature from its default returns an error. Even with Fable 5.2 above it, Anthropic recommends Opus 5.5 as the starting point for most coding and knowledge work, and at $4/$20 per 1M tokens it costs half as much as Fable 5.2.

  • Reasoning score: 83.9 (#5) · Overall 76.0
  • Reasoning controls: output_config.effort (low to max), medium by default, can't be turned off
  • Pricing: $4 input / $20 output per 1M tokens, with thinking billed as output
  • Context window: 1M tokens, 128K max output
  • Best for: Long-horizon agentic work and coding

6. GPT-6.1 Terra

OpenAI released GPT-6.1 Terra on October 26 as a lower-cost tier below Sol, and it ranks sixth on reasoning at 82.8, level with Kimi K3.1 Pro's estimated score and 7.0 points behind GPT-6.2 Astra at an eighth of the price. At $1/$5 per 1M tokens it's the cheapest model in the top 10, and OpenAI's model page lists no reasoning benchmark numbers for it.

Terra is also the only model in the top 10 that turns reasoning fully off. Its reasoning.effort runs from none to max, with medium as the default, so one model ID can serve extraction and classification steps with reasoning off and harder steps with it on. Tool calling requires the Responses API, and prompts over 272K input tokens bill at 2x input and 1.5x output, the same surcharge as GPT-6.2 Astra.

  • Reasoning score: 82.8 (#6) · Overall 69.1
  • Reasoning controls: reasoning.effort (none to max), medium by default, none turns reasoning off
  • Pricing: $1 input / $5 output per 1M tokens, with reasoning billed as output and the same surcharge above 272K input tokens as GPT-6.2 Astra
  • Context window: 1.05M tokens, 128K max output
  • Best for: High-volume steps and low-cost reasoning

7. Kimi K3.1 Pro

Kimi K3.1 Pro, Moonshot AI's October 23 update to Kimi K3, is the only open-weight model in the top 10, and it ranks seventh on reasoning at an estimated 82.8. The estimate reflects that Moonshot's release post gives no GPQA Diamond or Humanity's Last Exam figures in text, though Moonshot does report a hallucination rate of 8.3% on its own Kimi Hallu Bench Hard, down from 23.4% for K3.

Its reasoning_effort defaults to max, so an unset request runs it at its most expensive setting, and reasoning can't be turned off. Multi-turn tool loops need the complete assistant message passed back, with reasoning_content and tool_calls included. Moonshot published the weights on November 4 under its own license, and the hosted API costs the same $2/$10 as Sonnet 5.5, with a Highspeed variant at $4/$20 that Moonshot says runs two to three times faster. Running it yourself instead of paying per token is the tradeoff that Kimi K3.1 Pro vs Fable 5.2 weighs against Anthropic's top model.

  • Reasoning score: 82.8, estimated (#7) · Overall 71.5
  • Reasoning controls: reasoning_effort (low, high, max), max by default, can't be turned off
  • Pricing: $2 input / $10 output per 1M tokens, $4 / $20 for Highspeed
  • Context window: 1M tokens, 131K output by default and configurable up to 1M
  • Best for: Open-weight reasoning with a path to self-hosting

8. Muse Wonder Full

Muse Wonder Full is the larger of the two Muse Wonder models Meta announced October 12, and it ranks eighth on reasoning at an estimated 82.5. It has no API, though: Wonder Full is only available through Meta's Muse Code Power ($50 per month) and Maximum ($200 per month) plans.

For an app, the version you can call is Muse Wonder (muse-wonder-1.0) at $4/$20 per 1M tokens, which scores 73.0 on reasoning, 9.5 points below Wonder Full. If you leave reasoning_effort unset, Muse Wonder reasons at max, and setting none returns an HTTP 400 error. Raw reasoning stays private, with summaries available on the Responses API when you ask for them, though Meta notes a summary isn't guaranteed on every response.

  • Reasoning score: 82.5, estimated (#8) · Overall 71.0
  • Reasoning controls: No API; Muse Wonder uses reasoning.effort (minimal to max), max by default, can't be turned off
  • Pricing: No API, Muse Code subscription from $50 per month ($4 / $20 per 1M for Muse Wonder)
  • Context window: 1M tokens
  • Best for: Muse Code users, with Muse Wonder for API calls

9. Claude Sonnet 5.5

Claude Sonnet 5.5 ranks ninth on reasoning at 78.9, but its coding score of 78.8 is the highest on the Respan Index, so it earns its place here on different work than the models above it. Anthropic reports 64.5% on Humanity's Last Exam with tools, and the Opus 5.5 vs Sonnet 5.5 comparison covers when it can stand in for Opus.

Effort defaults to high on the Claude API, and Anthropic recalibrated the levels relative to Sonnet 5, so a setting carried over from the older model won't behave the same way. Thinking can't be disabled, but thinking: {"type": "between_tools"} limits it to the steps between tool calls at low, medium, or high effort. At xhigh or max that setting returns a 400 error, and while it's on, effort can't change partway through a conversation. At $2/$10 per 1M tokens it costs half of Opus 5.5 and the same as Kimi K3.1 Pro.

  • Reasoning score: 78.9 (#9) · Overall 73.4
  • Reasoning controls: output_config.effort (low to max), high by default, thinking can be limited to between tool calls
  • Pricing: $2 input / $10 output per 1M tokens, with thinking billed as output
  • Context window: 1M tokens, 128K max output
  • Best for: Coding and high-volume agent steps

10. Gemini 4 Argon

Google describes Gemini 4 Argon as its efficient frontier reasoning model, built to spend fewer thinking tokens per turn than Carbon, and it ranks tenth on reasoning at 78.5, 5.5 points behind Carbon at the same per-token price. Google's own results still put Argon ahead of Carbon on RiemannBench for math, at 76.0% against 72.0%.

Argon uses thinking_level with low, medium, and high, defaulting to medium, and setting minimal returns a 400 error. It can return up to 1M output tokens through long decode, which matters when one call has to generate a whole codebase or a long document. Its introductory $2/$10 rate runs through November 30 and then rises to $4/$20, at which point Neon costs less per token and outscores it on reasoning by 14.6 points, so Gemini 4 Hydrogen vs Argon vs Carbon is worth reading before you build around it.

  • Reasoning score: 78.5 (#10) · Overall 69.3
  • Reasoning controls: thinking_level (low, medium, high), medium by default, can't be turned off
  • Pricing: $2 input / $10 output per 1M tokens through November 30, 2026, then $4 / $20, with thinking billed as output
  • Context window: 1M tokens, up to 1M output with long decode
  • Best for: Long-horizon reasoning on a smaller thinking budget

What reasoning models actually cost

Standard rates per 1M tokens, from each provider's pricing page as of November 19, 2026:

ModelInputOutputPricing note
Gemini 4 Neon$3.00$15.00$0.15 cached input
Claude Fable 5.2$8.00$40.00$16 / $80 in fast mode
GPT-6.2 Astra$8.00$40.002x input, 1.5x output above 272K, fast mode 2x
Gemini 4 Carbon$2.00$10.00Rises to $4 / $20 on Dec 1
Claude Opus 5.5$4.00$20.00Flat rate across the 1M window
GPT-6.1 Terra$1.00$5.00Same surcharge above 272K
Kimi K3.1 Pro$2.00$10.00$4 / $20 for Highspeed
Muse Wonder$4.00$20.00API version of Muse Wonder Full
Claude Sonnet 5.5$2.00$10.00Half of Opus 5.5's rate
Gemini 4 Argon$2.00$10.00Rises to $4 / $20 on Dec 1

Reasoning models don't carry a separate price list. The same per-token rates you'd compare across the best LLM APIs apply, and the thinking lands in the output column, which is the more expensive side of each rate. Several billing details change the real number on top of that:

  • Max-effort defaults - Kimi K3.1 Pro and Muse Wonder run at max unless you set effort, so an unset request spends as many thinking tokens as each model allows.
  • Truncated calls that still bill - On OpenAI models, max_output_tokens covers reasoning and visible output together. If reasoning uses up the limit, the response comes back incomplete and you can be billed with no visible output, which is why OpenAI recommends reserving at least 25,000 tokens.
  • Long-context surcharges - GPT-6.2 Astra and GPT-6.1 Terra bill the whole request at higher rates above 272K input tokens, while Opus 5.5 keeps one rate across its context window.
  • Time-based pricing - Carbon's and Argon's rates double on December 1, 2026, and fast mode on Fable 5.2 and GPT-6.2 Astra doubles the rate for faster output.

Effort level is usually the fastest lever to pull, and it stacks with other LLM cost optimization strategies like caching and batch discounts.

How to evaluate reasoning models

Published scores narrow the field, but you settle the choice by running candidates against your own workload.

Test on your own traffic, not leaderboards

Leaderboards rank these models on composite scores, but the labs report different LLM benchmarks, and none of those benchmarks used your prompts or your tool schemas. A model that leads on a science benchmark can still struggle with your function definitions or your output format.

Pull a few hundred real requests from production, including ones that failed, and run them through each candidate at two or three effort levels. That gives you a dataset that reflects what your users actually send.

Measure cost per completed task

Per-token price doesn't tell you what a task costs, because a model that thinks longer or needs a retry spends more tokens to get the same result. OpenAI's own figures show how large the gap can get: at max effort, GPT-6.1 Sol completed Terminal-Bench Science tasks for $5.47 each, against $23.80 for GPT-6 Astra.

Divide total spend, retries and failed calls included, by the number of tasks that passed your checks. That number is what to compare across models and effort levels.

Track latency and thinking-token spend

Reasoning tokens arrive before the first visible token, so higher effort shows up as a longer time-to-first-token as well as a bigger bill. Log reasoning tokens separately from visible output where the API reports them; Muse Wonder returns them as reasoning_tokens, and Gemini as total_thought_tokens.

Hidden reasoning can make a response look cheaper than it was. Anthropic notes that omitting thinking text reduces latency but not cost, so a fast response on Opus 5.5 can still carry a full thinking bill.

Score output quality with evals

Once you know cost and latency at each effort level, the remaining question is whether the cheaper setting still gets the answer right. Define what a good output looks like with an LLM judge, a deterministic code check, or human review, then run each model and effort level against the same dataset and keep the lowest setting that clears the bar.

The broader methods for evaluating LLMs in production apply here too, with effort level treated as one more variable in the experiment. With Respan evals, the same evaluator that scores your dataset before you ship can run on sampled live traffic afterward, so a model update or prompt change that degrades quality shows up before users report it.

How to route every task to the right reasoning model

Respan AI Gateway traffic with requests, errors, latency, tokens, and cost

Choosing a reasoning model per task only pays off if you can see what each call costs and prove that the cheaper choice held up. Respan is the AI router that gives you one API key for 1,000+ models at up to 55% off per token, with a 0% fee and built-in observability and automated evals, so every request it routes is traced, priced, and scored in the same place. Route, observe, and evaluate every LLM call.

Through Respan's AI gateway, every major provider sits behind one API, and moving a step to a different model means changing a single word. Respan Router (beta) takes that further: set the model to span-router, and it picks the model for each task based on the request's context, the capabilities it needs, and cache-aware cost estimates, billed at the serving model's rate with no markup.

  • Cost per request and per customer - Every request is logged with its latency and cost, broken down by model, request, and end customer, so you can see which feature or customer drives your spend.
  • Hard spend limits - Set soft and hard caps per key, per customer, or org-wide. The soft cap warns you as spend approaches it, and the hard cap stops spend at the ceiling.
  • Fallbacks when a provider fails - Define a fallback chain once, and traffic moves to the next model the moment a provider errors or rate-limits.
  • Traces for multi-step agents - Each LLM call, tool run, and agent turn becomes a span in one trace with its own input, output, latency, and cost.
  • Evals that close the loop - Build datasets from production logs, run experiments that compare models or prompt versions against your evaluators, and click any score to land on the trace behind it.
  • Alerts before customers notice - Monitors watch cost, errors, latency, and tokens, and send an alert to Slack, email, or a webhook the moment one breaches.

Instead of reading logs after the fact, use Respan to run observability in production, know when production shifts, and act before it spreads.

FAQ

What is the best AI reasoning model?

Gemini 4 Neon has the highest Respan Index reasoning score at 93.1, ahead of Claude Fable 5.2 (90.0) and GPT-6.2 Astra (89.8). Neon costs $3/$15 per 1M tokens against $8/$40 for the other two, and the best model for a given step is usually the cheapest one that passes your evals, which is why Respan Router (beta) picks the model per task so you don't have to hardcode that choice.

Which AI reasoning model is best for coding?

Claude Sonnet 5.5 leads Respan's ranking of the best LLM for coding at 78.8, followed by Claude Opus 5.5 (75.9) and Claude Fable 5.2 (73.0). Run the candidates on tasks from your own repository before committing; Respan experiments let you compare models side by side against the same dataset.

What is the best open-weight reasoning model?

Kimi K3.1 Pro, the only open-weight model in the top 10, with an estimated reasoning score of 82.8. Moonshot published its weights on November 4 under its own license, but at 2.8T total parameters, running it yourself takes serious hardware, so the hosted API at $2/$10 per 1M tokens is the practical path for many teams, the same tradeoff behind most open-weight models.

What is the cheapest AI reasoning model?

Among the top 10, GPT-6.1 Terra at $1 input and $5 output per 1M tokens, followed by Kimi K3.1 Pro, Claude Sonnet 5.5, Gemini 4 Carbon, and Gemini 4 Argon at $2/$10 (until Carbon and Argon rise to $4/$20 on December 1), then Gemini 4 Neon at $3/$15. The lowest per-token rate doesn't always mean the lowest cost per task, since a model that needs higher effort or retries to pass can cost more overall. Respan breaks down cost by model, request, and customer, so you can see what each task costs in production.

Do AI reasoning models cost more than standard models?

Yes, per request. Thinking tokens bill as output on top of the visible answer, and Anthropic bills them even when the thinking text isn't returned. You can bring the cost down by lowering effort, turning reasoning off where a model allows it (GPT-6.1 Terra) or limiting it (Claude Sonnet 5.5's between_tools setting), or routing simple steps to a non-reasoning model, and Respan shows cost per request so you can see which calls drive your spend.

Related articles

Built for AI agents.
Break less.
Ship more.