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: start with Claude Sonnet 5.5 in a Claude workflow or GPT-6.1 Sol in Codex, and escalate complex, long agent runs to Claude Opus 5.5. At today's standard API prices, an illustrative agent day costs $6 on either daily driver and $12 on Opus. For cheap, tightly scoped volume, GPT-6 Luna is the lowest-priced option here; for self-hosted terminal work, start with GLM-5.3.

Verified October 5, 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.

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.

Since September 22, 2026, Claude Opus 5.5 has replaced the older Opus recommendation; Claude Sonnet 5.5 followed on September 28, and GPT-6.1 Sol entered Codex on September 29. Move saved GPT-5.5 Codex defaults before its October 14, 2026 retirement from Codex with ChatGPT sign-in. That retirement does not apply to the OpenAI API. Anthropic release notes, Codex changelog, OpenAI retirement guidance.

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

Use Sonnet or Sol for everyday edits and reviews, Opus for sprawling work that needs sustained judgment, and a cheaper model for tasks whose correctness you can check automatically. 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. The agent-day column assumes 2 million uncached input tokens plus 200,000 billable output tokens across requests below the long-context price thresholds. Self-hosting has an infrastructure bill rather than a universal token list price.

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. Every source in this table was read on October 5, 2026. “Not published” means the primary pages checked did not provide a result for that exact model and benchmark. It does not mean a zero score. The terminal column uses Terminal-Bench 4.0 throughout.

Sonnet's 70.6% is above Opus's 66.4% on Anthropic's published terminal results. 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 evaluation and positioning, Opus evaluation notes.

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.

At $4 input / $20 output, the example day is $12, down from $15 on Opus 4.8 at the same token totals. 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. Current and legacy pricing, 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.

Claude Fable 5.1 remains an escalation option at $10 / $50, or $30 for the example day. Anthropic recommends it when demanding work still defeats Opus at higher effort. Do not make it the default because the tier sounds more capable: that is $18 extra per example day over Opus, and its premium must buy accepted work. Anthropic model selection and pricing.

GPT-6.1 Sol: The OpenAI Daily Driver

GPT-6.1 Sol is the practical OpenAI choice for recurring coding work: $2 / $10, or $6 for the example day. OpenAI positions it for complex coding at lower cost than Astra. 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.

The normalized bill falls from $16 on GPT-5.5 to $6 on Sol, a 62.5% reduction at identical token totals. That is price arithmetic, not a promise that both models use the same tokens to finish a ticket. For an API agent, use the Responses API for tool calling. GPT-5.5 list rate, Sol pricing and API support.

Sol's 1,050,000-token context has a price threshold: requests above 272K input tokens cost $4 / $15 for the whole request. If every request in an equivalent daily token total crosses that threshold, the model-token bill is $11, rather than $6. Its $0.10 cached-input rate can help repeated context, but a cache-dependent budget must use actual cache hits. OpenAI model pricing.

GPT-6 Luna is the volume option at $0.10 / $0.50, or $0.30 for the same illustrative day. Start with bounded jobs such as extracting fields from tickets or drafting commit messages, then check the outputs before routing more work to it. GPT-6 Astra costs $10 / $50, or $30 per example day; reserve it for jobs where Sol's accepted results justify paying more. Luna, Astra.

Gemini 3.8 Flash and 3.1 Pro Preview: Price Versus Autonomy

Gemini 3.8 Flash is Google's budget coding candidate at today's published rate. Its $2.25 example day makes it worth evaluating for repeated, testable changes. Its 1,048,576-token input limit also supports large inputs. Google pricing, Flash model limits.

Its limit for this decision is the published 19.1% Terminal-Bench 4.0 result. A low token rate alone does not make it the best unattended coding agent. 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. Google's own evaluation.

Gemini 3.5 Flash-Lite is a separate low-cost candidate for parsing and simple subagent work at $0.30 / $2.50, or $1.10 per example day. Gemini 3.1 Pro Preview remains available, but retain its Preview label. It costs $2 / $12 at or below 200K input tokens per prompt, rising to $4 / $18 above that threshold. The equivalent example-day totals are $6.40 and $11.60 respectively. Current Google prices, Flash-Lite intended use, Pro availability.

Pro is worth keeping if your existing workflow produces better accepted outcomes with it. It no longer wins this comparison merely by having a large context window: Flash, Sonnet and Sol also accept large contexts, with different prices and capabilities. Gemini 4 Argon is not an available recommendation.

Claude Sonnet 5.5 and Haiku 4.5: 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. It costs $2 / $10, has a 1M-token context, and produces a $6 example day. The same workload at Sonnet 4.6's $3 / $15 costs $9. The new price removes much of the reason to keep the older daily recommendation. Anthropic pricing, 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 4.5 remains the inexpensive Claude option at $1 / $5, with 200K context and a $3 example day. Use it for bounded automation if staying inside the Claude stack simplifies your system. It is more expensive at these token totals than Luna or Flash-Lite, so it is not the cheapest model overall. Anthropic's announced Haiku 5.5 is still forthcoming. Current Claude models, Sonnet launch and forthcoming Haiku.

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 the official Terminal-Bench 4.0 leaderboard records 41.8% with Claude Code at max effort. Evaluate it on a repository where commands, test runs and multi-step fixes dominate the work. Maker model card, official result.

If your existing stack uses DeepSeek, compare its accepted results before replacing it. Kimi K3 is also a serious candidate when your coding workflow needs image input. Moonshot releases its weights with a model-specific 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.

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 cards do not give a universal self-host token price, and a vendor's hosted API quote would 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. In an OpenAI workflow, compare Sol on the same jobs before paying for Astra. Keep Fable available when Opus still falls short, rather than routing every overnight job to the most expensive tier.

Daily driver: choose Sonnet 5.5 or GPT-6.1 Sol according to the tool you already use. Both cost $6 in the same uncached example. The choice flips when one produces more accepted patches with less review and retry work on your repository.

Cheap high volume: start by evaluating GPT-6 Luna for focused tasks. Compare Gemini 3.5 Flash-Lite if your Google workflow matters, or Haiku 4.5 if Claude integration 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. 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 because of the retirement date. Skip a blanket Fable or Astra default without evidence that the extra $24 over a $6 example day buys useful work. Skip a model ranking built by copying older versions' scores or comparing different Terminal-Bench generations.

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.

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. Both cost $2 input and $10 output per million tokens at the base Standard API rate. For complex, long Claude agent runs, escalate to Opus 5.5. The cited primary pages do not establish a current SWE-bench tie, and Sol's availability in Work 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 complex long agent runs, GPT-6 Luna for focused cheap volume, and GLM-5.3 as the self-host starting point for terminal-heavy work. Gemini 3.8 Flash is a lower-cost Google candidate at its current promotional rate. Choose using accepted results in the tool you will actually run.

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. These are open-weight choices with model-specific terms to check, not a promise of free inference. You still pay for infrastructure and operations; the model cards and the official results cited above do not establish one winner across all coding jobs.

Which model is cheapest for coding?

GPT-6 Luna has the lowest base text-token prices in this lineup: $0.10 input and $0.50 output per million tokens. The 2-million-input, 200,000-output example costs $0.30. Gemini 3.5 Flash-Lite costs $1.10 and Haiku 4.5 costs $3 at those totals. 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. 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.