Best AI Model for Coding in 2026 (Ranked by SWE-bench and Price)

The best AI model for coding in 2026 depends on the job: Claude Sonnet 5.5 is the first daily coding and review pick in a Claude workflow, GPT-6.1 Sol is the Codex default to evaluate, and Claude Opus 5.5 is the first escalation for difficult agent runs and high-risk reviews. Kimi K3.1 Pro is the new hosted challenger for long agent work; GPT-6.1 Terra is the lower-cost step down for bounded coding and visual fixes. For cheap bulk work, compare GPT-6 Luna with Claude Haiku 5.5; for self-hosted terminal work, start with GLM-5.3. Gemini 4 Carbon is announced, not usable yet.

Verified October 29, 2026. Prices, availability, model limits and the benchmark sources below were rechecked against the makers' live pages and the benchmarks' own leaderboards on this date.

Since the October 14 verification, Moonshot released Kimi K3.1 Pro on October 23, and OpenAI released GPT-6.1 Terra on October 26 with same-day Codex and GitHub Copilot access; both enter the recommendations below, while K3.1 Pro's downloadable weights remain promised for November 4. Moonshot announcement, OpenAI announcement, Copilot announcement.

A model is the intelligence that generates the answer. A coding tool or agent, such as Claude Code, Cursor or Codex, supplies the repository, runs commands and returns test results. Choose a tool that supports your workflow, then select the model it actually exposes.

Best AI Model for Coding in 2026: The Verdict, in One Table

The first six rows are the ranked buying shortlist: Sonnet and Sol for daily work, Opus for difficult work, K3.1 Pro as the long-run value challenger, Terra for bounded tasks, and Gemini 4 Argon for a Google workflow. The remaining usable models are specialist or existing-integration choices; Carbon is an unranked announcement to watch. This is a buying order based on workflow fit, published coding evidence and cost, not a single benchmark leaderboard. The premium tiers earn their place only when your repository's results justify the extra spend.

The table uses USD per million tokens, input then output, at the published Standard API rate; Carbon’s row is announced future pricing. The agent-day column assumes 2 million base-rate input tokens plus 200,000 billable output tokens, spread across short requests. It is a model-token subtotal, not a subscription quote or a measured cost per completed ticket.

Where These Models Are Available

Codex: OpenAI documents native access to GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Astra and GPT-6 Luna. Kimi K3 and K3.1 Pro also have a maker-documented route through a custom Kimi provider in Codex CLI and desktop, using separate Kimi credentials. That route is distinct from an OpenAI subscription model. No native Codex listing for the other models in this comparison was established by the pages checked. Codex model guidance, Kimi's Codex setup.

Claude Code: Sonnet 5.5, Opus 5.5, Fable 5.1 and Haiku 5.5 are documented choices; Haiku 4.5 remains available on documented provider routes. Pin the full model name because aliases differ across cloud providers and gateways. Z.ai documents a custom endpoint for GLM-5.3, and Moonshot documents one for Kimi K3 and K3.1 Pro. These use the outside provider's access and billing. The checked Claude Code page does not establish native OpenAI or Gemini model choices. Claude Code configuration, Z.ai integration, Kimi integration.

Cursor: its current pricing catalog lists Sonnet 5.5, Opus 5.5, Fable 5.1, Haiku 5.5, Gemini 4 Argon, Gemini 3.8 Flash and Gemini 3.1 Pro. Third-party access requires Premium or above and prepaid usage; team fees can add to model rates. The page mentions Gemini 3.5 Flash-Lite in fee exclusions, but does not establish its current picker availability. It does not confirm the four GPT-6 models above, Haiku 4.5, GLM-5.3, Kimi K3 or K3.1 Pro in the catalog read for this check. Treat those as unconfirmed, not interchangeable with older GPT names or GLM-5.3-Flash. Cursor models and pricing.

GitHub Copilot: the supported-model list includes all four OpenAI models above, all five Claude models in the table, Gemini 4 Argon, Gemini 3.8 Flash and Kimi K3. K3.1 Pro, GLM-5.3 and Gemini 3.5 Flash-Lite are not listed. Gemini 3.1 Pro was retired from Copilot on September 1, 2026. Plan, client and administrator controls still apply; Fable has additional access conditions. Terra's launch covers paid Copilot plans, including CLI and cloud agent, with gradual rollout. Supported models and retirement dates, Terra rollout.

If you would rather not pick a model, GitHub Copilot Fusion, generally available since October 28, 2026, routes and coordinates eligible models for each task: select Fusion in the model picker, then choose an effort level. GitHub’s announcement.

Why "Best on SWE-bench" Does Not Settle the Decision

An agent benchmark measures a model inside a particular tool setup, with a particular effort setting and task budget. It does not establish that model's success rate on your backlog. SWE-bench Verified uses a human-filtered set of 500 repository issues; Terminal-Bench evaluates agents working through terminal tasks. SWE-bench, Terminal-Bench.

Here is the attributable evidence for this lineup, read on October 29, 2026. “Not found” means the primary pages checked did not establish a result for that exact model and benchmark. It does not mean a zero score. The terminal column uses Terminal-Bench 4.0 throughout; maker evaluations and the benchmark's public leaderboard remain separately labeled.

Sonnet's launch result is above Opus's on Anthropic's published terminal evaluation. That is a reason to try the cheaper model first, not proof that Sonnet handles every long migration better. Anthropic still describes Opus as stronger on complex, open-ended work requiring sustained judgment. Opus's reported evaluation also used production safeguards and documented fallback models. Sonnet positioning, Opus evaluation notes.

The official leaderboard separately lists 61.8% for Sonnet 5.5 and 64.8% for Opus 5.5, both in Claude Code at max effort. Those are different runs from the launch results above. Fable and Astra also have distinct maker and public-leaderboard results, both labeled in the table. Terra and K3.1 Pro's table entries come from their makers, not entries on that public leaderboard. Moonshot's comparison takes peers' maximum reported results from vendor posts or the leaderboard, so it is not one uniformly rerun comparison. Terminal-Bench leaderboard, Moonshot methodology.

Do not mix benchmark versions. Moonshot publishes 88.3% on Terminal-Bench 2.1 for Kimi K3, while Z.ai publishes 28.3% on Terminal-Bench 3.0 for GLM-5.3. Neither number can be ranked against the 4.0 column above. Google's Pro card separately publishes 54.2% on SWE-bench Pro Public; that is another test, not a substitute for a missing terminal result. Kimi card, GLM card, Gemini Pro card.

Claude Opus 5.5: The Escalation for Long Agent Runs

Claude Opus 5.5 is the starting escalation for a migration, an ambiguous bug across services, or an audit where the agent must keep checking its own assumptions. The recommendation follows Anthropic's stated intended use and published evaluations.

Its 1M-token context and 128K maximum output support larger work, but they do not make filling the window a good default. Give an overnight agent a defined change, acceptance tests and a spend limit. Model limits.

The practical migration limit matters: Opus 5.5's thinking cannot be disabled, and integrations that force a particular tool must follow the new tool-choice rules. Update the client before switching an automated job's model string. Opus migration behavior.

For code review, use Sonnet first on ordinary pull requests and Opus when a change crosses service boundaries or touches authorization, billing or data integrity. Ask for the failing path and a reproducible test for each finding. This is a review-routing recommendation, not a measured claim that every Opus review beats every Sonnet review.

Claude Fable 5.1 remains an escalation option when demanding work still defeats Opus at higher effort. Do not make it the default because the tier sounds more capable: its premium must buy accepted work. Anthropic model selection.

GPT-6.1 Sol and Terra: The OpenAI Daily Driver and Step Down

GPT-6.1 Sol is the practical OpenAI choice for recurring coding work. OpenAI positions it for complex coding at lower cost than Astra. Its launch evaluation reports Sol matching Astra on DeepSWE v1.1 at roughly one-fifth the cost, a maker comparison separate from the terminal leaderboard. OpenAI’s coding evaluation. It is available in Codex, ChatGPT Work and the API; check your plan and client rather than assuming it appears in ordinary Chat. Model page, availability.

GPT-6.1 Terra is the first step down to evaluate when a Sol task is well scoped. A screenshot-backed layout fix, a bounded bug with a failing test, or a routine patch is a better first assignment than an unattended cross-service migration. OpenAI's launch reports a substantial terminal-coding gain over the previous Terra, but explicitly keeps Sol and Astra ahead for the most complex coding. That makes Terra a useful budget agent, not a universal new number one. OpenAI's evaluation and guidance.

Terra is also a sensible candidate for checking a UI against a supplied screenshot. Its documented strengths include visual reasoning and computer use. Do not turn those strengths into an unsupported claim that it is the best security reviewer. Terra model page.

Sol and Terra both expose 1,050,000-token API context windows, but Sol defaults to 500,000 tokens in Codex and Work. Tool calling requires the Responses API. Terra requires Codex CLI 0.159.0 or later; select the exact gpt-6.1-terra model after updating. API limits and subscription behavior are separate. Sol specifications, Terra specifications, Codex model selection.

GPT-6 Luna is the volume option for bounded jobs such as extracting fields from tickets, drafting commit messages or making mechanical edits checked by tests. OpenAI reports 66.6% on DeepSWE v1.1 at max effort for Luna; that is a different evaluation from its terminal score above. OpenAI’s Luna evaluation. GPT-6 Astra remains the premium escalation; reserve it for jobs where its accepted results justify paying more than Sol. Luna, Astra.

Gemini 4 Carbon and Argon, 3.8 Flash and 3.1 Pro Preview: Price Versus Autonomy

Google announced Gemini 4 Carbon on October 28, 2026. Its status is announced, not usable yet for general developers. Access starts with selected cyber defenders and trusted Fairwind partners. Google reports internal engineering runs lasting days to find and patch subtle codebase bugs, and more than 2× Argon’s success rate for generating secure code on internal benchmarks. These are Google’s claims. Carbon stays unranked until public access permits evaluation; its table prices are announced future rates, with no launch date or introductory expiry stated. Google’s Carbon announcement.

Gemini 4 Argon is the first Google frontier model to evaluate for complex coding, now sixth in this buying shortlist. K3.1 Pro adds a hosted long-run challenger, while Terra adds a cheaper bounded-agent option. Neither addition makes Argon irrelevant to a Google-based team.

Argon is generally available in the paid Gemini API as gemini-4-argon. Antigravity's current model page also allows it on paid Google AI Pro, eligible premium-device trials, Google AI Ultra and Enterprise. Gemini CLI's own docs say it has not been optimized for Argon and direct users to Antigravity. Google API pricing and availability, Antigravity model access, Gemini CLI guidance.

Google publishes 77.9% on DeepSWE v1.1, 55.0% on FrontierSWE v2 and 91.9% on Vibe Code Bench for Argon. Its comparison card lists Opus 5.5 at 74.2%, 62.3% and 90.3%, respectively. Argon leads on DeepSWE and Vibe Code Bench, while Opus leads on FrontierSWE and the terminal result in the table. None of these is a SWE-bench Verified score. Google's model card; the rounded external scores agree with FrontierSWE and Vibe Code Bench.

These are Google's published results, not tests performed for this article. Google computed Argon's DeepSWE result with a mini-swe agent setup and its Terminal-Bench result itself; it sourced FrontierSWE and Vibe Code Bench from their official leaderboards. Different agent setups and budgets limit direct comparisons. Google's evaluation methodology.

For a Google-based team evaluating a repository migration, Argon belongs in the trial. Keep a working daily driver until your own tickets show a gain, and include the scheduled end of introductory pricing in the buying decision.

Gemini 3.8 Flash is Google's budget coding candidate. Its 1,048,576-token input limit supports large inputs, but its published terminal result does not make it the first unattended-agent pick. Flash model limits.

For a founder generating predictable patches with tests, compare the accepted-patch bill. For an open-ended migration, keep Opus or Sol in the evaluation. A low token rate alone does not establish low total cost.

Gemini 3.5 Flash-Lite is a separate candidate for parsing and simple subagent work. Gemini 3.1 Pro Preview remains an API option, but retain its Preview label and account for its long-prompt price band. Pro is worth keeping if your existing workflow produces better accepted outcomes with it; its continued API availability does not reverse its Copilot retirement. Flash-Lite intended use, Pro API availability.

Claude Sonnet 5.5 and Haiku: Daily Coding and Cheap Claude Tasks

Claude Sonnet 5.5 is the default to evaluate for routine features, refactors and reviews in a Claude workflow. Its 1M-token context supports repository work, but a clear change request and relevant files still matter more than filling the window. Model limits.

For a CTO choosing a default, the important question is whether Sonnet finishes the ticket with acceptable review effort. Escalate an unresolved failure to Opus after preserving the failing test and a clear explanation. Repeatedly asking the cheaper model to retry can erase its price advantage; repeated blind escalation can waste the premium.

Claude Haiku 5.5 is the current Claude option for high-volume, latency-sensitive work such as extraction, routing and subagents. It ties Luna's base input and output prices for short prompts, so Luna is no longer the sole cheapest token-rate choice in this comparison. Haiku's price rises above its short-prompt band. Evaluate it for mechanical edits with tests; do not infer long-agent performance from its low rate. Haiku specifications and intended use.

Claude Haiku 4.5 remains a legacy option for existing integrations. Keep it where the integration and accepted results justify it, rather than presenting it as the current budget Claude default. Its older coding score is not a score for Haiku 5.5. Current and legacy model pricing.

Kimi K3.1 Pro: A Hosted Challenger for Long Agent Work

Kimi K3.1 Pro ranks fourth in the buying shortlist because it makes a serious long-run evaluation affordable at daily-driver token rates. Moonshot's terminal result is close to its cited Opus result, while the price table places it alongside Sonnet and Sol. That is enough to put it in a migration trial, not enough to declare it the winner on your repository.

Give it an ambiguous bug with acceptance tests or a multi-file refactor that currently needs escalation. Compare completed work, review effort and retries against Opus. For routine code review, Sonnet remains the first recommendation; the sources checked do not establish a separate K3.1 Pro review advantage.

The API model is kimi-k3.1-pro and supports a 1,048,576-token context. Kimi Code uses the distinct k3.1-pro identifier and provides documented Codex and Claude Code integrations. Its Plus tier allows 262,144 tokens for this model; Pro and above unlock the full window. Those subscription limits are separate from the pay-as-you-go API. API specifications, Kimi Code model configuration.

Keep thinking enabled and start a fresh session when switching. Moonshot warns that a mistyped coding-plan model ID can silently route to K2.8 Preview without thinking. Record the resolved model before comparing results. Do not substitute K3.1 Pro for K3 in a self-host plan: the announced weight release is still ahead. Kimi Code routing rules, Weight-release commitment.

The Open-Weight Models Worth Self-Hosting

GLM-5.3 is the self-host starting point here for terminal-heavy coding. Z.ai's own model card documents local serving through vLLM and SGLang, and its official terminal result is listed above. Evaluate it on a repository where commands, test runs and multi-step fixes dominate the work. Maker model card.

If your existing stack uses DeepSeek, compare its accepted results before replacing it. Kimi K3, the older released model, is also a candidate when your coding workflow needs image input. Moonshot releases its weights under the Kimi K3 License and documents local deployment. Its published results use different evaluations, so this page does not declare a universal open-weight winner. Moonshot's own card.

K3.1 Pro's promised release does not make its weights downloadable today. For a deployment that must stay on your infrastructure now, compare GLM-5.3 and K3; revisit Pro when the actual files and license are available. Moonshot release commitment.

Self-hosting means serving the weights on infrastructure you control. That can suit a privacy-bound operator, but it does not mean free inference. Calculate allocated GPU-hours × your GPU-hour cost, then add storage, power and operations. The hosted prices in the table let you budget an API evaluation; they do not price your own fleet.

For a team that already operates inference, compare cost per accepted ticket and the capacity needed at peak load. For a solo builder without that infrastructure, start with the hosted daily driver. The ownership decision should precede the benchmark decision.

Which Model Should You Actually Pick

Long agent runs: use Opus 5.5 as the first escalation for complex, open-ended work. Evaluate K3.1 Pro as the lower-priced hosted challenger. In an OpenAI workflow, compare Sol on the same jobs before paying for Astra. Keep Fable available when Opus still falls short.

Code review: start with Sonnet 5.5 for ordinary pull requests, or Sol if Codex is already your review environment. Escalate high-risk, cross-service changes to Opus. Require concrete evidence for findings; a terminal benchmark is not a review-accuracy test.

Daily driver: choose Sonnet 5.5 or GPT-6.1 Sol according to the tool you already use. Try Terra when the task is bounded or a screenshot is central to the fix. The choice flips when one produces more accepted patches with less review and retry work on your repository.

Google frontier coding: keep Carbon on the watchlist until general access opens; evaluate Argon in the paid Gemini API, a supported Antigravity plan, Cursor or Copilot. Switch only if accepted patches and review effort justify it, including after its introductory rate ends.

Cheap high volume: compare GPT-6 Luna and Haiku 5.5 for short, focused jobs. Terra is the next OpenAI candidate when Luna's mistakes make the cheapest rate unhelpful. Compare Flash-Lite if your Google workflow matters. Use automated checks to decide what can stay on the cheap tier.

Self-host: evaluate GLM-5.3 for terminal work and Kimi K3 for workflows that need image input. K3.1 Pro is a hosted option pending its weight release. Check the model license, memory capacity and actual serving performance before substituting an infrastructure estimate for the API bill.

Skip GPT-5.5 as a new Codex default with ChatGPT sign-in: it retired there on October 14, 2026. API access and Codex sessions using an API key are unaffected. Skip a blanket Fable or Astra default without evidence that the premium buys useful work. Skip a model ranking built by copying older versions' scores or comparing different Terminal-Bench generations. OpenAI retirement guidance.

OpenRouter offers a common API endpoint and model catalog. If you use it for the comparison, confirm the exact model is listed and record any fallback route; do not assume every fresh release is available through the same provider. OpenRouter documentation.

The next move is concrete: take a bug fix, a feature and a refactor from your backlog. Give each candidate the same acceptance tests and spend limit. Record accepted patches, total token charges, retries and human review time. Change the default on those results, and remove the retired model from saved configurations that use ChatGPT sign-in.

The tool still controls how the model sees your repository and which commands it can run. Use the guide to the best AI coding agents, the Codex vs Claude Code vs Cursor comparison, and the Grok Build, Claude Code and Codex comparison when choosing that layer.

Is ChatGPT or Claude better at coding in 2026?

For daily coding, compare Claude Sonnet 5.5 with GPT-6.1 Sol in Codex or ChatGPT Work. For difficult long runs and high-risk reviews, Opus 5.5 is the first escalation here. Terra is a lower-cost candidate for bounded coding, not the recommended replacement for Sol on the hardest work. Work and Codex availability does not imply availability in ordinary Chat.

What is the best AI model for coding right now?

Sonnet 5.5 or GPT-6.1 Sol for a daily driver, Opus 5.5 for difficult long runs, and Kimi K3.1 Pro as the hosted value challenger. Terra suits bounded coding and visual fixes; Argon is the Google frontier choice. Luna and Haiku 5.5 cover focused cheap volume, while GLM-5.3 is the self-host starting point for terminal-heavy work.

What is the best free or open-source model for coding?

Start with GLM-5.3 for self-hosted terminal coding and evaluate Kimi K3 when image input matters. K3.1 Pro's weights are promised, not yet a verified deployment option. Check model-specific terms and budget for infrastructure and operations; the cited evidence does not establish one winner across all coding jobs.

Which model is cheapest for coding?

GPT-6 Luna and short-prompt Claude Haiku 5.5 tie for the lowest base text-token rates in this lineup. Haiku's rate rises above its short-prompt band. Use the pricing table for the exact rates and thresholds; the cheapest accepted patch can differ if a model needs more retries or review.

Does the model or the coding tool matter more?

Choose the tool first, then a model it supports. Claude Code, Cursor and Codex determine the repository context, tool permissions and testing loop; the model determines how it reasons within that setup. Native model-picker access and a custom provider configuration are different routes. Compare the complete setup on your own tickets rather than buying a model solely for a leaderboard number.

Want the current model-and-tool picks without re-reading a leaderboard every month? Join the newsletter and get the shortlist, with prices and benchmarks, the week they change.