There is no single best LLM for coding, and a universal ranking usually mixes benchmarks that don't measure the same thing. What you can compare honestly is what each model costs on a realistic coding-agent workload, how much context it documents, whether it supports tool calls and thinking, and what its vendor has actually published. This guide does that for every coding-capable model ATP serves.
Coding-capable models
Vendors
Lowest cost per agent task
Documented ≥1M context
Reading order, not a ranking. Each pick states when it fits and the vendor claim behind it; prices are ATP list rates.
Pick it if you want one default model for agentic coding in Claude Code-style harnesses.
Anthropic calls it its most agentic Sonnet — strong across planning, implementation, debugging and large refactors — and reports it closing much of the gap to Opus 4.8 on agentic and coding evaluations.
Anthropic — Introducing Claude Sonnet 5Pick it if your agents run on OpenAI / Codex tooling and tokens per task matter.
OpenAI says GPT-5.5 matches GPT-5.4's per-token latency while using significantly fewer tokens on the same Codex tasks.
OpenAI — Introducing GPT-5.5Pick it if you need a low-cost loop model for tool-heavy agent steps.
Google pitches Gemini 3 Flash as Pro-grade reasoning at Flash latency and cost, and reports 78% on SWE-bench Verified.
Google — Introducing Gemini 3 FlashPick it if you want an open-weight coding specialist for long plan–edit–run–debug loops.
Moonshot built it on K2.6 for long-horizon software engineering, with thinking always on, and reports 62.0 on Kimi Code Bench v2 versus 50.9 for K2.6.
Moonshot AI — Kimi-K2.7-Code model card (Hugging Face)Pick it if you hand an agent a whole codebase and need a long window on an open-weight model.
Z.ai positions GLM-5.2 for project-level codebase takeover and long-horizon refactoring over a 1M-token context, and reports 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro.
Z.ai developer docs — GLM-5.2 overviewPick it if you need editor completion and bulk code tasks at a very low rate.
DeepSeek supports fill-in-the-middle completion in non-thinking mode and documents a 1M-token context for the V4 line.
DeepSeek API — models & pricingPick it if a task is hard enough that a failed run costs more than the rate premium — use it as the escalation tier.
Anthropic reports gains in agentic coding and ships dynamic workflows that let Claude Code dispatch and verify parallel subagents, over a documented 1M-token context.
Anthropic — Introducing Claude Opus 4.8Every coding-capable model on ATP, priced on the same coding-agent workload at list rates. The bar compares monthly cost without caching.
Monthly cost at current prices
| Model | Price | Context | Per task | Per month | With 80% cache hits | Per month · Relative · log scale |
|---|---|---|---|---|---|---|
| Qwen 3.7 Flashqwen-3-7-flash | $0.03 / $0.13 | — | $0.02 | $22 | $10 | |
| Qwen Plus Characterqwen-plus-character | $0.115 / $0.287 | — | $0.08 | $78 | $33 | |
| Qwen 3.6 Flashqwen-3-6-flash | $0.165 / $0.99 | — | $0.13 | $129 | — | |
| DeepSeek V4 Flashdeepseek-v4-flash | $0.20 / $0.40 | 1M | $0.13 | $132 | $55 | |
| Qwen 3.6 Plusqwen-3-6-plus | $0.276 / $1.651 | — | $0.22 | $215 | — | |
| Gemini 3.5 Flash Litegemini-3-5-flash-lite | $0.30 / $2.50 | 1.05M | $0.26 | $255 | $125 | |
| Qwen 3.7 Plusqwen-3-7-plus | $0.40 / $1.60 | — | $0.29 | $288 | $134 | |
| Gemini 3 Flash (Preview)gemini-3-flash-preview | $0.50 / $3.00 | — | $0.39 | $390 | — | |
| DeepSeek V3.2deepseek-v3-2 | $0.57 / $1.71 | — | $0.39 | $393 | $174 | |
| Kimi K2.6kimi-k2.6 | $0.95 / $4.00 | 262K | $0.69 | $690 | — | |
| Kimi K2.7 Codekimi-k2.7-code | $0.95 / $4.00 | — | $0.69 | $690 | $325 | |
| GLM-5glm-5 | $1.00 / $3.20 | 200K | $0.70 | $696 | — | |
| Claude Haiku 4.5claude-haiku-4-5 | $1.00 / $5.00 | 200K | $0.75 | $750 | $318 | |
| GPT-5.6 Lunagpt-5.6-luna | $1.00 / $6.00 | — | $0.78 | $780 | — | |
| GLM-5.2glm-5-2 | $1.40 / $4.40 | — | $0.97 | $972 | $434 | |
| Qwen 3.6 Maxqwen-3-6-max | $1.29 / $7.71 | — | $1.01 | $1,005 | — | |
| GLM-5.1glm-5-1 | $1.54 / $4.84 | — | $1.07 | $1,069 | — | |
| Gemini 3.6 Flashgemini-3-6-flash | $1.50 / $7.50 | — | $1.13 | $1,125 | — | |
| Gemini 3.5 Flashgemini-3-5-flash | $1.50 / $9.00 | 1.05M | $1.17 | $1,170 | — | |
| Qwen 3.8 Maxqwen-3-8-max | $2.00 / $6.00 | — | $1.38 | $1,380 | $540 | |
| Gemini 3.1 Pro (Preview)gemini-3-1-pro-preview | $2.00 / $12.00 | — | $1.56 | $1,560 | — | |
| DeepSeek V4 Prodeepseek-v4-pro | $2.40 / $4.80 | 1M | $1.58 | $1,584 | $528 | |
| Qwen 3.7 Maxqwen-3-7-max | $2.50 / $7.50 | — | $1.73 | $1,725 | $765 | |
| GPT-5.6 Terragpt-5.6-terra | $2.50 / $15.00 | — | $1.95 | $1,950 | — | |
| GPT-5.4gpt-5.4 | $2.50 / $15.00 | 1.05M | $1.95 | $1,950 | — | |
| Claude Sonnet 5claude-sonnet-5 | $3.00 / $15.00 | 1M | $2.25 | $2,250 | $954 | |
| Claude Sonnet 4.6claude-sonnet-4-6 | $3.00 / $15.00 | — | $2.25 | $2,250 | $954 | |
| Kimi K3kimi-k3 | $3.00 / $15.00 | — | $2.25 | $2,250 | $954 | |
| Claude Opus 5claude-opus-5 | $5.00 / $25.00 | 1M | $3.75 | $3,750 | $1,590 | |
| Claude Opus 4.8claude-opus-4-8 | $5.00 / $25.00 | 1M | $3.75 | $3,750 | $1,590 | |
| Claude Opus 4.7claude-opus-4-7 | $5.00 / $25.00 | 1M | $3.75 | $3,750 | $1,590 | |
| GPT-5.6 Solgpt-5.6-sol | $5.00 / $30.00 | — | $3.90 | $3,900 | — | |
| GPT-5.5gpt-5.5 | $5.00 / $30.00 | 1.05M | $3.90 | $3,900 | — | |
| Claude Fable 5claude-fable-5 | $10.00 / $50.00 | — | $7.50 | $7,500 | $3,180 | |
| GPT-5.4 Progpt-5.4-pro | $30.00 / $180.00 | 1.05M | $23.40 | $23,400 | — |
USD list price, estimate only. “With 80% cache hits” bills 80% of input at the model's cache-read rate and is shown only where ATP lists one (cache-write charges not included). Tiered models use the band for 20K-token requests. Real agents vary a lot by task, harness and retries — measure yours from the per-request logs.
What vendors have published about coding, with sources. These numbers come from different suites, versions and harnesses, run by the vendors themselves — they are not comparable with each other, and we don't rank on them.
Vendor-reported coding evidence
| Model | Vendor claim | Reported by | Source |
|---|---|---|---|
| Gemini 3 Flash (Preview)gemini-3-flash-preview | SWE-bench Verified 78% | Google — Introducing Gemini 3 Flash | |
| Kimi K2.6kimi-k2.6 | SWE-bench Verified 80.2%, SWE-bench Pro 58.6% | Moonshot AI | Moonshot AI — Kimi-K2.6 model card (Hugging Face) |
| Kimi K2.7 Codekimi-k2.7-code | Kimi Code Bench v2 62.0 (K2.6: 50.9) | Moonshot AI | Moonshot AI — Kimi-K2.7-Code model card (Hugging Face) |
| GLM-5.2glm-5-2 | Terminal-Bench 2.1 81.0, SWE-bench Pro 62.1 | Z.ai | Z.ai developer docs — GLM-5.2 overview |
| GLM-5.1glm-5-1 | SWE-bench Pro 58.4 | Z.ai | Z.ai developer docs — GLM-5.1 overview |
| Qwen 3.6 Maxqwen-3-6-max | First place on six coding / agent benchmarks incl. SWE-bench Pro and Terminal-Bench 2.0 (scores not given in our source) | Alibaba (via Decrypt) | Decrypt — Alibaba drops Qwen 3.6 Max Preview |
| Claude Sonnet 5claude-sonnet-5 | Closes much of the gap to Opus 4.8 on agentic and coding evaluations | Anthropic | Anthropic — Introducing Claude Sonnet 5 |
| Claude Haiku 4.5claude-haiku-4-5 | Matches Claude Sonnet 4's coding performance at one-third the cost and more than twice the speed (Oct 2025) | Anthropic | Anthropic — Introducing Claude Haiku 4.5 |
| GPT-5.5gpt-5.5 | Significantly fewer tokens than GPT-5.4 on the same Codex tasks, at matching per-token latency | OpenAI | OpenAI — Introducing GPT-5.5 |
Coding agents resend repository context, loop through tools and retry, so tokens per finished change run far above a chat reply. Treat them like a product surface: their own project, key, allowlist and cap.
# a dedicated project + key for coding agents
export ANTHROPIC_BASE_URL="https://api.atptoken.ai"
export ANTHROPIC_AUTH_TOKEN="atp-•••"
export ANTHROPIC_API_KEY=""
export ANTHROPIC_MODEL="claude-sonnet-5" # must be on the project allowlist
claude
# a model outside the allowlist → 403; spend stops at the project allocationFAQ
Rates, capabilities and context on this page come from the ATP model catalog. Other claims cite the sources below.
Issue a project key, allowlist the models you want to test, and compare them on your own traffic.