Best AI Model for Coding: Quick Answer (June 2026)
The best AI model for coding in June 2026 is Claude Opus 4.8 for most teams: 88.6% on SWE-bench Verified at $5/$25 per million tokens, and the highest SWE-bench Pro score (69.2%) of any model you can actually buy. The higher-scoring Claude Fable 5 (95.0%) is suspended under a June 12 export-control directive. GPT-5.5 ties Opus 4.8 on SWE-bench Verified at 88.7%, and Claude Haiku 4.5 is the cheapest per solved task at about $0.13 of output per benchmark point.
This page divides the two columns every other ranking keeps apart: SWE-bench Pro scores from Scale's standardized leaderboard, official per-token prices, and the output-dollar cost per solved benchmark point. Updated June 28, 2026: every score and price refreshed to current models, GPT-5.6 (limited preview), GPT-5.5, GLM-5.2, and Claude Opus 4.8 added, and Claude Fable 5 marked suspended.
Best buyable model
Claude Opus 4.8
- 88.6% SWE-bench Verified
- 69.2% SWE-bench Pro (vendor)
- $5 / $25 per M tokens, 1M context
- Fable 5 (95.0%) is suspended, see note
Best on a standardized harness
GPT-5.4
- 59.10% SWE-bench Pro (Scale SEAL, #1)
- $2.50 / $15 per M tokens
- Superseded by GPT-5.5 ($5/$30), still sold
Cheapest per solved point
Claude Haiku 4.5
- 39.45% SWE-bench Pro (Scale SEAL)
- $1 / $5 per M tokens
- ~$0.13 output per Pro point
GPT-5.6 (Sol, Terra, Luna) previewed June 26 as a limited, government-gated preview, not on the public API. GLM-5.2 (744B MoE / 40B active, June 13) shipped with a usable 1M context at $1.40/$4.40. Claude Fable 5 was suspended June 12 under a US export-control directive. MiniMax M3 (June 1) is open weights at 80.5% SWE-bench Verified. Claude Opus 4.8 (May 28, 88.6%) is the current Anthropic default at $5/$25.
12 Models Ranked: SWE-bench Pro x Price x Cost per Solved Point
On a standardized harness, gpt-5.4 solves the most SWE-bench Pro tasks (59.10%) and Claude Haiku 4.5 solves them for the least money (about $0.13 of output per point). The table below uses Scale's SEAL public-set leaderboard, which runs every model through the same scaffolding on SWE-bench Pro (1,865 tasks, 41 professional repositories, scored Pass@1). The last column divides official output price by score: dollars of output tokens per benchmark point solved. Lower is more cost-effective.
| Model | SWE-bench Pro | $/M input / output | Output $ per Pro point |
|---|---|---|---|
| gpt-5.4 (xHigh) | 59.10% | $2.50 / $15 | $0.25 |
| Muse Spark (Meta) | 55.00% | not published | n/a |
| Claude Opus 4.6 (thinking) | 51.90% | $5 / $25 | $0.48 |
| Gemini 3.1 Pro (thinking) | 46.10% | $2 / $12 (≤200K tokens) | $0.26 |
| Gemini 3 Pro (preview) | 43.30% | preview | n/a |
| gpt-5.2-codex | 41.04% | superseded by gpt-5.5 | n/a |
| Claude Haiku 4.5 | 39.45% | $1 / $5 | $0.13 |
| Qwen3 Coder 480B (open weights) | 38.70% | self-host | n/a |
| Gemini 3 Flash | 34.63% | see Gemini pricing | n/a |
| Kimi K2 Instruct (open weights) | 27.67% | self-host | n/a |
Three things fall out of the combined view. GPT-5.4 wins on score and is competitive on cost per point at $0.25. Haiku 4.5 solves 67% as many tasks as gpt-5.4 at a third of its output price, making it the cost-per-point leader at roughly $0.13. And Opus 4.6, the top Claude entry Scale has tested, pays a 2x cost-per-point premium over gpt-5.4 for 7.2 fewer points on this harness, which is exactly why Anthropic publishes its own numbers (covered below).
Scale also runs a private (commercial) set drawn from proprietary startup codebases, 276 instances the models have never seen. Models that top the public set do not automatically top unseen code: Muse Spark, for example, scores 55.00% on the public set but 44.70% on the private set. If your repo looks nothing like open-source Python, weight the private-set ordering, not the public leaderboard.
SWE-bench Verified Leaderboard (June 2026)
On SWE-bench Verified, Claude Fable 5 leads at 95.0% but is suspended; the top buyable models are GPT-5.5 (88.7%) and Claude Opus 4.8 (88.6%), a near tie. SWE-bench Verified is older, Python-only, and partially contaminated, but it is still the number every launch post quotes. Every score below is vendor self-reported; the llm-stats tracker lists 0 of them as independently verified. The June 2026 top ten:
SWE-bench Verified: Top 10 (June 2026)
Source: llm-stats.com tracker, June 2026. Vendor self-reported; higher = more GitHub issues resolved.
Claude holds 4 of the top 5; GPT-5.5 edges Opus 4.8 by 0.1 point. The 80-percent cluster is mostly open weights.
Two structural facts in this chart. First, the Claude line pulled away at the very top: Fable 5's 95.0% is 6.4 points above Opus 4.8 and 14.4 above the 80-percent cluster, but Fable 5 is unavailable, so the live ceiling is the GPT-5.5 / Opus 4.8 tie near 88.6%. Second, the 80-percent cluster is now mostly open weights. DeepSeek-V4-Pro-Max (80.6%) ties Gemini 3.1 Pro exactly, and you can download its MIT-licensed weights.
Which Claude Model Is Best for Coding?
Claude Opus 4.8 (claude-opus-4-8) is the best Claude model for coding: 88.6% SWE-bench Verified, 69.2% SWE-bench Pro on Anthropic's harness, $5/$25 per million tokens, 1M context with no long-context surcharge. Anthropic ships five current coding-relevant models. Exact API IDs, prices, and the decision logic:
| Model (API ID) | Coding benchmarks | $/M in / out | Context / max output |
|---|---|---|---|
| Claude Fable 5 (claude-fable-5) | 95.0% SWE-bench Verified, 80.0% SWE-bench Pro (vendor); suspended | $10 / $50 | 1M / 128K |
| Claude Opus 4.8 (claude-opus-4-8) | 88.6% Verified, 69.2% Pro (vendor) | $5 / $25 | 1M / 128K |
| Claude Opus 4.7 (claude-opus-4-7) | 87.6% Verified | $5 / $25 | 1M / 128K |
| Claude Sonnet 4.6 (claude-sonnet-4-6) | 79.6% SWE-bench Verified | $3 / $15 | 1M / 64K |
| Claude Haiku 4.5 (claude-haiku-4-5) | 39.45% SWE-bench Pro (Scale SEAL) | $1 / $5 | 200K / 64K |
Default: Opus 4.8
claude-opus-4-8 at $5/$25 is the working default for coding agents: 88.6% SWE-bench Verified, 69.2% SWE-bench Pro on Anthropic's harness (the highest of any buyable model), and 1M context with no long-context surcharge. A fast-mode research preview is priced at $10/$50, about 3x cheaper than fast mode on Opus 4.6/4.7.
Ceiling: Fable 5 (currently suspended, see note)
claude-fable-5 ($10/$50, GA June 9, 2026) adds 6.4 points of SWE-bench Verified and 10.8 points of vendor SWE-bench Pro over Opus 4.8. Suspended June 12, 2026 per a US export-control directive that bars access by any foreign national. While suspended, Opus 4.8 is the practical ceiling pick.
Volume: Sonnet 4.6
claude-sonnet-4-6 at $3/$15 carries a 1M context window and scores 79.6% SWE-bench Verified. Use it for high-throughput agent loops where Opus pricing compounds: CI review bots, test generation, batch transforms. Batch API halves it to $1.50/$7.50.
Quick edits and subagents: Haiku 4.5
claude-haiku-4-5 at $1/$5 is the cost-per-point leader on Scale's leaderboard (~$0.13 of output per Pro point). Route single-file edits, lint fixes, and explore-style subagents here; cache hits cost $0.10/M.
Treat Claude Sonnet 4, Opus 4, and Opus 4.1 as legacy and migrate to claude-sonnet-4-6 / claude-opus-4-8. Note the tokenizer change too: Opus 4.7 and later (including Fable 5) can produce up to 35% more tokens for the same text than pre-4.7 models, so compare per-request costs, not just per-token rates. Full price tables on the Anthropic API pricing page.
One model to set aside: Claude Mythos 5 (93.9% SWE-bench Verified as Mythos Preview) is a limited-availability model restricted to approved Project Glasswing partners. It was suspended alongside Fable 5 on June 12, 2026 under the same directive. There is no self-serve access and no active availability, so it is not a practical coding pick.
Claude Opus 4.8 vs GPT-5.5: The $5-Input Flagships
Opus 4.8 and GPT-5.5 both cost $5/M input and tie on SWE-bench Verified (88.6% vs 88.7%); Opus 4.8 wins repo-scale software engineering on SWE-bench Pro (69.2% vs 58.6%) and costs less output ($25 vs $30), while GPT-5.5 is the model OpenAI now ships through Codex. The splits:
| Dimension | Claude Opus 4.8 | GPT-5.5 |
|---|---|---|
| SWE-bench Verified (vendor) | 88.6% | 88.7% |
| SWE-bench Pro (vendor) | 69.2% | 58.6% |
| Pricing ($/M in / out) | $5 / $25 | $5 / $30 |
| Cached input ($/M) | $0.50 (cache hit) | $0.50 |
| Context window | 1M | ~1M (128K max output) |
Head-to-Head: The Race Card
GPT-5.3 Codex vs Claude Opus 4.6 across 7 dimensions
Scores based on benchmarks, developer surveys, and hands-on testing as of February 2026. Neither model "wins" overall — it depends on your workflow.
Opus 4.8 wins repo-scale software engineering (69.2% vs 58.6% SWE-bench Pro on vendor harnesses) and is cheaper on output ($25 vs $30 per M). GPT-5.5 edges SWE-bench Verified by 0.1 point and is what OpenAI now points Codex at: the dedicated -codex variants are retired, so the general gpt-5.5 is the current Codex model. If your work lives in a CLI agent, Codex pricing changes the math; if it lives in long-horizon repo edits, Opus 4.8 does.
OpenAI previewed GPT-5.6 on June 26, 2026 in three variants: Sol (flagship), Terra (balanced), Luna (fast). It is gated, available only to about 20 government-approved companies after a US government request, not on the public API, with general availability promised in the coming weeks. OpenAI published no SWE-bench score at preview; a secondary tracker lists GPT-5.6 Sol at 88.8% on Terminal-Bench 2.1. Until it ships broadly, GPT-5.5 ($5/$30) is the OpenAI coding model you can actually buy.
Open-Source Models: 80% SWE-bench Verified at a Tenth of the Price
The best open-source coding model in June 2026 is DeepSeek-V4-Pro-Max at 80.6% SWE-bench Verified, tied with Gemini 3.1 Pro and self-hostable under MIT. MiniMax M3 (June 1, 2026) matches it at 80.5% with a 1M context window, and Kimi K2.6 reaches 80.2%. Official API prices and self-host terms:
| Model | SWE-bench Verified | $/M in / out (official API) | Context / license |
|---|---|---|---|
| DeepSeek-V4-Pro-Max (1.6T / 49B active) | 80.6% (top open weights) | $0.435 / $0.87 (V4 Pro) | 1M / MIT |
| DeepSeek V4 Flash (284B / 13B active) | 79.0% (Flash-Max) | $0.14 / $0.28 | 1M / MIT |
| morph-dsv4flash (DeepSeek V4 Flash on Morph) | bf16 activations, codegen spec decoding + kernels | $0.139 / $0.278 | MIT weights, hosted |
| MiniMax M3 | 80.5% | $0.60 / $2.40 | ~1M / open weights |
| Kimi K2.6 (1T / 32B active) | 80.2% | $0.95 / $4.00 | 256K / open weights |
| Qwen3.6 Plus | 78.8% | $0.50 / $3.00 | 1M / Qwen |
| GLM-5.2 (744B / 40B active) | no official SWE-bench; 62.1% Pro (third-party) | $1.40 / $4.40 | 1M / open weights |
| morph-glm52-744b (GLM-5.2 on Morph) | bf16 activations, codegen kernels | $1.1 / $4.1 | 1M, hosted |
The arithmetic that matters: MiniMax M3 produces output at $2.40/M against Opus 4.8's $25/M, a 10.4x gap, while trailing it by 8.1 points on SWE-bench Verified (80.5 vs 88.6). DeepSeek V4 Flash sets the absolute floor at $0.28/M output with a 1M-token context. The price gap is mostly model size, not provider margin, which is the core fact of how AI inference is priced. For teams with data-sovereignty requirements, DeepSeek V4's MIT license means the 80.6% model is self-hostable outright. Deeper coverage on the best open-source coding model page.
Open weights are identical everywhere, the serving stack is not. Most serverless providers quantize activations to fp8 to cut cost, which degrades output quality. Morph serves DeepSeek V4 and GLM-5.2 with 16-bit (bf16) activations and does not quantize them, so output matches the reference weights. That makes Morph the place to run open models when fidelity matters. For coding specifically, Morph adds speculative decoding tuned on code plus custom low-level inference kernels built for code generation. morph-dsv4flash (DeepSeek V4 Flash) runs at $0.139/M input and $0.278/M output, and morph-glm52-744b (GLM-5.2) at $1.1/M input and $4.1/M output. See the full catalog on Morph models and pricing.
Best AI Model for Coding at $0
The best free path to real coding capability in June 2026 is DeepSeek V4's MIT-licensed open weights (self-host) or its near-free hosted API at $0.14/$0.28. Four no-credit-card options:
| Option | What you get | Limit |
|---|---|---|
| Codex CLI on ChatGPT Free | Codex CLI with GPT-5.5, $0/mo | Lowest usage limits; Plus ($20/mo) raises them |
| Z.AI free GLM tier | Free Flash-tier GLM models on the Z.AI API | Flash tier only; GLM-5.2 flagship is paid at $1.40/$4.40 |
| Qwen on Alibaba Model Studio | Free token allowance for new users across Qwen models | Time-limited trial |
| DeepSeek V4 open weights | MIT-licensed weights, self-host V4 Flash (284B/13B active) | Your GPU cost; hosted API is $0.14/$0.28 anyway |
The honest framing: free tiers are for evaluation and light use. DeepSeek V4 Flash's paid API at $0.14/M input ($0.0036/M on cache hits) and $0.28/M output is close enough to zero that most teams skip self-hosting unless data cannot leave their network.
Why Vendor Scores Run 20 Points Above Scale's Leaderboard
Anthropic reports Fable 5 at 80.0% on SWE-bench Pro. Scale's standardized leaderboard tops out at 59.10% (gpt-5.4). Both numbers are real. The difference is the harness: Scale runs every model through identical scaffolding; vendors run their own tuned agent stacks. (Vendor SWE-bench Verified numbers are also self-reported; llm-stats lists 0 as independently verified.)
Same Benchmark, Different Harness: SWE-bench Pro
Vendor-reported (Anthropic scaffold) vs Scale SEAL standardized scaffolding.
The vendor-vs-standardized gap is 17-21 points for the same model families. The harness is the variable.
The practical conclusion has not changed since 2025: the scaffold around the model accounts for more variance than swapping frontier models. Before paying a 2x token premium, fix retrieval, context management, and tool design. Subagent architecture and context engineering move scores more than model choice does.
A mid-tier model in a strong harness beats a frontier model in a weak one. Tools like WarpGrep (semantic codebase search for terminal agents, $0 for 100k requests) upgrade the harness for every model you route through it.
Per-Task Routing: Which Model for Which Job
The most cost-effective setups in June 2026 route by task, not by loyalty: send the hard 20% to Opus 4.8 and the cheap 80% to Haiku 4.5 or DeepSeek V4 Flash. Numbers-backed defaults:
| Task | Route to | Why (verified numbers) |
|---|---|---|
| Overnight refactor, 50+ files | Claude Opus 4.8 | 69.2% SWE-bench Pro (vendor), 1M context, no long-context surcharge |
| Hardest debugging / migration runs | Claude Opus 4.8 (Fable 5 suspended) | 88.6% Verified, 69.2% Pro, the highest of any buyable model |
| Quick edits, lint fixes, subagents | Claude Haiku 4.5 | $1/$5, ~$0.13 output per Pro point, $0.10/M cache hits |
| Terminal / Codex workflows | GPT-5.5 | $5/$30, the model OpenAI now ships through Codex |
| Standardized-harness ceiling | gpt-5.4 | 59.10% Scale SEAL SWE-bench Pro at $2.50/$15 |
| High-volume batch / CI bots | DeepSeek V4 Flash or MiniMax M3 | $0.28/M and $2.40/M output, both ~1M context |
| Budget proprietary, long prompts | Gemini 3.1 Pro | 46.10% Scale Pro at $2/$12 (≤200K); input doubles to $4 above 200K |
| Data sovereignty / self-host | DeepSeek V4 (MIT) | 80.6% SWE-bench Verified (Pro Max), weights on Hugging Face |
| Codebase search for any agent | WarpGrep + any model | Model-agnostic retrieval; $0 for 100k requests |
Cost levers that apply across routes: Anthropic's Batch API is 50% off input and output, prompt-cache reads are 0.1x base input, and DeepSeek cache hits drop input to $0.0036/M. A routing setup that pins 80% of traffic to Haiku 4.5 or DeepSeek V4 Flash and reserves Opus 4.8 for the hard 20% typically beats any single-model subscription. Doing the split automatically needs a classifier: Morph's Router scores each prompt by difficulty and domain in ~180ms and returns the cheapest capable model, so a coding agent gets cheaper and faster at once (see the model lineup and pricing). Claude Code Router makes that per-request routing concrete inside the terminal agent, and Claude Code models covers harness-side defaults.
Frequently Asked Questions
What is the best AI model for coding in June 2026?
Claude Opus 4.8 (claude-opus-4-8) is the best AI model for coding you can actually buy: 88.6% SWE-bench Verified and 69.2% SWE-bench Pro (vendor) at $5/$25 per million tokens, with a 1M context window. The higher-scoring Claude Fable 5 (95.0% SWE-bench Verified, $10/$50) is suspended as of June 12, 2026 (see note above). GPT-5.5 ties Opus 4.8 on SWE-bench Verified at 88.7%. On Scale's standardized SWE-bench Pro leaderboard, gpt-5.4 leads at 59.10%, ahead of Muse Spark (55.00%) and Claude Opus 4.6 (51.90%). Cost-adjusted, Claude Haiku 4.5 ($1/$5) is the cheapest per solved benchmark point at roughly $0.13 of output.
What is the best LLM for coding?
"Best LLM for coding" and "best AI model for coding" are the same question. By raw capability, the top buyable LLM in June 2026 is Claude Opus 4.8 (88.6% SWE-bench Verified, $5/$25); Claude Fable 5 scored 95.0% but is suspended (see note above). The best open-weights LLM is DeepSeek-V4-Pro-Max at 80.6%, with MiniMax M3 (80.5%) and Qwen3.7 Max (80.4%) within 0.2 points. The cheapest per solved point is Claude Haiku 4.5 at about $0.13. For most teams the best answer is not one LLM but a router that sends easy work to a cheap or open model and reserves a frontier model for hard edits.
Which Claude model is best for coding?
Claude Opus 4.8 (API ID claude-opus-4-8) is the default and the practical ceiling: 88.6% SWE-bench Verified, 69.2% SWE-bench Pro on Anthropic's harness (the highest of any buyable model), $5/$25, 1M context with no long-context surcharge. Claude Fable 5 (claude-fable-5, $10/$50) scored 95.0% Verified but is suspended (see note above). Claude Sonnet 4.6 (claude-sonnet-4-6, $3/$15, 79.6% Verified) is the volume pick with a 1M context, and Claude Haiku 4.5 (claude-haiku-4-5, $1/$5) handles quick edits and subagents. Treat Sonnet 4, Opus 4, and Opus 4.1 as legacy and migrate off them.
What is the best Codex model for coding in 2026?
OpenAI now points Codex at the general gpt-5.5 model; there is no dedicated gpt-5.5-codex variant, and gpt-5.3-codex is deprecated. So the best Codex model in June 2026 is gpt-5.5 ($5/$30 per million tokens, 88.7% SWE-bench Verified vendor-reported, 58.6% SWE-bench Pro). GPT-5.6 (Sol, Terra, Luna) previewed June 26, 2026 but is a limited preview gated to about 20 government-approved companies and not yet on the public API; OpenAI did not publish a SWE-bench score for it at launch.
What are the SWE-bench Pro scores for coding models in 2026?
Scale SEAL public set (standardized scaffolding, June 2026): gpt-5.4 xHigh 59.10%, Muse Spark 55.00%, Opus 4.6 thinking 51.90%, Gemini 3.1 Pro thinking 46.10%, Gemini 3 Pro 43.30%, gpt-5.2-codex 41.04%, Haiku 4.5 39.45%, Qwen3 Coder 480B 38.70%, Gemini 3 Flash 34.63%, Kimi K2 Instruct 27.67%. Vendor-reported numbers run higher: Anthropic reports Fable 5 at 80.0% and Opus 4.8 at 69.2% on its own scaffold. All vendor numbers are self-reported; llm-stats lists 0 as independently verified.
What is the best free AI model for coding?
Four real $0 paths in June 2026: Codex CLI is included with a ChatGPT Free sign-in (lowest usage limits); Z.AI offers free Flash-tier GLM models on its API; Alibaba Model Studio gives new users a free token allowance across Qwen models; and DeepSeek V4's MIT-licensed weights are self-hostable. DeepSeek V4 Flash's paid API is near-free at $0.14/M input, $0.28/M output.
What is the best open-source AI model for coding?
DeepSeek-V4-Pro-Max leads open weights at 80.6% SWE-bench Verified, tied with Gemini 3.1 Pro. MiniMax M3 scores 80.5% (1M context, $0.60/$2.40), Kimi K2.6 80.2%, and Qwen3.6 Plus 78.8%. DeepSeek V4 ships under MIT: V4 Pro (1.6T total / 49B active) costs $0.435/$0.87 on the official API, V4 Flash (284B/13B) costs $0.14/$0.28 with a 1M context. GLM-5.2 (744B MoE / 40B active) shipped without an official SWE-bench number; third-party trackers put it at 62.1% on SWE-bench Pro.
How much do the top coding models cost per million tokens?
Output price ladder, June 2026: DeepSeek V4 Flash $0.28, MiniMax M2.7 $0.72, MiniMax M3 $2.40, Qwen3.6 Plus $3.00, Qwen3.7 Max $3.75, Kimi K2.6 $4.00, GLM-5.2 $4.40, Gemini 3.1 Pro $12, gpt-5.3-codex $14, gpt-5.4 and Claude Sonnet 4.6 $15, Claude Opus 4.8 $25, GPT-5.5 $30, Claude Fable 5 $50 (suspended). Inputs range from $0.14/M (DeepSeek V4 Flash) to $10/M (Fable 5).
Why do vendor benchmark scores differ from Scale's leaderboard?
Scale runs every model through identical standardized scaffolding on SWE-bench Pro's 1,865 tasks across 41 repositories, scored Pass@1; vendors run their own tuned harnesses. The same model family scores 51.90% (Opus 4.6 on Scale) versus 69.2% (Opus 4.8 on Anthropic's harness). That 17-to-21-point spread is the harness, which is why agent tooling moves results more than model swaps.
Which AI model is most cost-effective for coding in 2026?
Dividing output price by Scale SEAL SWE-bench Pro score: Claude Haiku 4.5 about $0.13 of output per point, gpt-5.4 $0.25, Gemini 3.1 Pro $0.26, Claude Opus 4.6 $0.48. For raw per-token cost with a 1M context, DeepSeek V4 Flash at $0.14/$0.28 is the floor.
Sources
Primary sources behind the scores and prices on this page (June 28, 2026):
- Scale SEAL SWE-bench Pro public leaderboard (standardized harness scores)
- llm-stats SWE-bench Verified tracker and SWE-bench Pro aggregate (vendor self-reported)
- SWE-bench Pro paper (1,865 tasks / 41 repos, benchmark definition)
- Anthropic Claude pricing and Fable / Mythos access notice (suspension)
- OpenAI API pricing, Codex models, and GPT-5.6 preview
- Google Gemini API pricing
- DeepSeek API pricing and Z.ai GLM-5.2 docs
- Meta Muse Spark announcement
Stop Debating Models. Start Searching Codebases.
WarpGrep adds semantic codebase search to any terminal agent. Works with Opus 4.8, GPT-5.5, Gemini 3.1 Pro, DeepSeek V4, GLM-5.2, or any model. $0 for 100k requests, $1 per 1M on Pro. The harness matters more than the model.