Compare the models behind coding agents by release date and API price, with links to the providers’ documentation.
Recent additions, newest first. API prices in USD per million tokens; each card names its tier and checked date. Context tiers, caching, and tool fees can change the bill. Agent subscriptions are priced separately.
· Meta
$4 / $20
Input / output per 1M tokens · Standard API; cache reads $0.20; no long-context premium
Pricing checked October 12, 2026
Meta’s new model for complex coding and extended agent work. Available through the API as muse-wonder-1.0 and in Muse Code.
· Meta
Subscription only
Muse Code Power Usage or Maximum Usage; no public API token rate
Pricing checked October 12, 2026
The higher-capability Wonder variant, available exclusively in Muse Code subscriptions. It is a separate model, not Wonder with max reasoning.
· JetBrains
Self-hosted
Infrastructure costs apply; no hosted API rate is listed here.
Pricing checked October 12, 2026
A compact 12B mixture-of-experts model with 2.5B active parameters, trained for repository exploration, edits, and checking fixes. Suitable for local coding agents and subagents.
· Anthropic
$0.1 / $0.5
Input / output per 1M tokens · Prompts up to 100K tokens; cache reads $0.01. Above 100K: $0.50 input / $2.50 output, $0.05 cache reads per million tokens.
Pricing checked October 12, 2026
A small, fast model for repeated coding tasks and subagents working alongside Sonnet or Opus. Supports adjustable reasoning effort.
· Mistral AI
$0.68 / $2.09
Input / output per 1M tokens · Public-preview sale price; cache reads $0.07. Standard rate: $1.36 input / $4.18 output, $0.14 cached input per million tokens.
Pricing checked October 12, 2026
Mistral’s new multimodal model for coding and agent tasks, with a 1M-token context window.
· OpenAI
$2 / $10
Input / output per 1M tokens · standard up to 272K input; cache reads $0.10; above 272K: $4 input / $15 output
Pricing checked October 2, 2026
The updated Sol model for coding and computer use, with a 1.05M-token context window. Use the Responses API for tool calling.
· Anthropic
$2 / $10
Input / output per 1M tokens · standard API; cache reads $0.10
Pricing checked October 12, 2026
A faster, lower-cost alternative to Opus 5.5, with a 1M-token context window.
· Xiaomi
$0.435 / $0.87
Input / output per 1M tokens · Xiaomi overseas real-time API; cache reads $0.0036 per million tokens. Batch and ultraspeed service rates differ.
Pricing checked October 12, 2026
Xiaomi’s flagship multimodal model for complex coding projects and long-running agent tasks, with 1M context.
· Xiaomi
$0.14 / $0.28
Input / output per 1M tokens · Xiaomi overseas real-time API; cache reads $0.0028 per million tokens. Batch and ultraspeed service rates differ.
Pricing checked October 12, 2026
The lower-cost MiMo-V2.6 model for frequent coding-agent calls and large workloads, with native multimodal input and 1M context.
· Anthropic
$4 / $20
Input / output per 1M tokens · standard API; cache reads $0.20
Pricing checked October 2, 2026
The newest Opus release. Standard input and output tokens cost 20% less than Opus 5.
· OpenAI
$2 / $10
Input / output per 1M tokens · standard; input up to 272K
Pricing checked October 2, 2026
The original GPT-6 Sol release. GPT-6.1 Sol is the newer version at the same standard input and output rates, with a lower cache-read price.
· OpenAI
$0.1 / $0.5
Input / output per 1M tokens · standard; input up to 272K
Pricing checked October 2, 2026
The lower-cost GPT-6 option for routine work and high-volume agent tasks.
· xAI
$2 / $6
Input / output per 1M tokens · standard; prompts below 200K
Pricing checked October 2, 2026
A 500K-context model with configurable reasoning. Its faster variant is available through Cursor and Grok Build.
· DeepSeek
$0.3 / $1.2
Input / output per 1M tokens · peak, uncached; off-peak $0.15 input / $0.60 output; cache hits $0.006 peak / $0.003 off-peak
Pricing checked October 2, 2026
An open-weights model with native image input and 1M context. The deepseek-flash API alias serves V4.1; legacy V4 Flash IDs also route here.
· OpenAI
$10 / $50
Input / output per 1M tokens · standard; input up to 272K
Pricing checked October 2, 2026
OpenAI’s most capable broadly available model, combining coding, computer use, and research in long-running tasks.
· Meta
$1.25 / $4.25
Input / output per 1M tokens · Standard API; cache reads $0.15. Contributor: $0.10 input / $0.20 output, $0.002 cached input; content may train Meta models
Pricing checked October 12, 2026
Updates Spark’s coding and agent behavior, with max reasoning in Standard mode. Available through Meta Model API and Muse Code.
$0.75 / $3.75
Input / output per 1M tokens · standard promo through Dec 31, 2026; $1.50 input / $7.50 output from Jan 1, 2027
Pricing checked October 2, 2026
Google’s latest Flash model for software engineering and agent workflows. The separate Cyber variant has restricted access.
· Anthropic
$10 / $50
Input / output per 1M tokens · standard API
Pricing checked October 2, 2026
Generally available for coding and knowledge work. The related Mythos 5.1 model is limited to trusted-access programs.
· Alibaba
Self-hosted
Infrastructure or third-party inference costs apply; hosted Qwen3.8 Flash is a related production variant.
Pricing checked October 12, 2026
An open-weight model for repository coding and long-running tool use. Its sparse architecture targets efficient inference over long contexts.
· Alibaba
Self-hosted
Infrastructure or third-party inference costs apply; hosted provider rates vary.
Pricing checked October 12, 2026
A dense 27B multimodal model for coding and agent work, with configurable thinking and a smaller deployment footprint than large mixture-of-experts models.
· Meta
Self-hosted
No Meta-hosted API rate; infrastructure or third-party hosting costs apply
Pricing checked October 12, 2026
A 30B dense text-and-image model distilled from Spark for local agents and coding. Apache 2.0 weights; 128K default context.
Newest releases appear first. Every price has a checked or snapshot date. Turn on historical columns to see the August 26, 2026 benchmark snapshot and dated usage figures. An unscored model has no comparable result in this snapshot; it is not a zero score.
Looking for typed choices, scores, or classification? See the Decision Models Index.
| Model | Weights | Agent examples | ||
|---|---|---|---|---|
| Muse WonderNew releaseMetaMeta’s new model for complex coding and extended agent work. Available through the API as muse-wonder-1.0 and in Muse Code.Wonder launchModel availabilityAPI pricing | $20 output$4 input / 1M tokensChecked October 12, 2026Standard API; cache reads $0.20; no long-context premium | Oct 12, 2026 | Closed | Muse Code |
| Muse Wonder FullNew releaseMetaThe higher-capability Wonder variant, available exclusively in Muse Code subscriptions. It is a separate model, not Wonder with max reasoning.Requires Power Usage or Maximum Usage. Not available through the pay-as-you-go API.Wonder launchModel availabilityMuse Code subscriptions | Subscription onlyChecked October 12, 2026Muse Code Power Usage or Maximum Usage; no public API token rate | Oct 12, 2026 | Closed | Muse Code |
| Mellum2.1New releaseJetBrainsA compact 12B mixture-of-experts model with 2.5B active parameters, trained for repository exploration, edits, and checking fixes. Suitable for local coding agents and subagents.Downloadable Apache-2.0 weights. Provider benchmark claims are not mixed into the historical coding score.Mellum2.1 releaseModel card and weights | Self-hostedChecked October 12, 2026Infrastructure costs apply; no hosted API rate is listed here. | Oct 8, 2026 | Open | Local coding-agent runtimes |
| Claude Haiku 5.5New releaseAnthropicA small, fast model for repeated coding tasks and subagents working alongside Sonnet or Opus. Supports adjustable reasoning effort.Haiku 5.5 releaseAPI pricing | $0.5 output$0.1 input / 1M tokensChecked October 12, 2026Prompts up to 100K tokens; cache reads $0.01. Above 100K: $0.50 input / $2.50 output, $0.05 cache reads per million tokens. | Oct 7, 2026 | Closed | Claude-compatible coding agents |
| Mistral Large 4New releaseMistral AIMistral’s new multimodal model for coding and agent tasks, with a 1M-token context window.API public preview. Downloadable weights are announced for later in October; not yet available as of October 12.Large 4 announcementModel and preview pricing | $2.09 output$0.68 input / 1M tokensChecked October 12, 2026Public-preview sale price; cache reads $0.07. Standard rate: $1.36 input / $4.18 output, $0.14 cached input per million tokens. | Oct 6, 2026 | Partial | Mistral Vibe Code |
| GPT-6.1 SolNew releaseOpenAIThe updated Sol model for coding and computer use, with a 1.05M-token context window. Use the Responses API for tool calling.Model and pricingRelease log | $10 output$2 input / 1M tokensChecked October 2, 2026standard up to 272K input; cache reads $0.10; above 272K: $4 input / $15 output | Sep 29, 2026 | Closed | Codex |
| Claude Sonnet 5.5New releaseAnthropicA faster, lower-cost alternative to Opus 5.5, with a 1M-token context window.Current API pricingModel, release date, and pricing | $10 output$2 input / 1M tokensChecked October 12, 2026standard API; cache reads $0.10 | Sep 28, 2026 | Closed | Claude Code |
| MiMo-V2.6 ProNew releaseXiaomiXiaomi’s flagship multimodal model for complex coding projects and long-running agent tasks, with 1M context.MIT-licensed weights are available. The API uses updated checkpoints that address repeated tool calls; the downloadable fixes have the MOPD suffix.Series releaseAPI pricingUpdated weightsTool-call update | $0.87 output$0.435 input / 1M tokensChecked October 12, 2026Xiaomi overseas real-time API; cache reads $0.0036 per million tokens. Batch and ultraspeed service rates differ. | Sep 22, 2026 | Open | MiMo Code / compatible agent runtimes |
| MiMo-V2.6 FlashNew releaseXiaomiThe lower-cost MiMo-V2.6 model for frequent coding-agent calls and large workloads, with native multimodal input and 1M context.MIT-licensed weights are available. The API uses updated checkpoints that address repeated tool calls; the downloadable fixes have the MOPD suffix.Series releaseAPI pricingUpdated weightsTool-call update | $0.28 output$0.14 input / 1M tokensChecked October 12, 2026Xiaomi overseas real-time API; cache reads $0.0028 per million tokens. Batch and ultraspeed service rates differ. | Sep 22, 2026 | Open | MiMo Code / compatible agent runtimes |
| Claude Opus 5.5New releaseAnthropicThe newest Opus release. Standard input and output tokens cost 20% less than Opus 5.Release and pricing | $20 output$4 input / 1M tokensChecked October 2, 2026standard API; cache reads $0.20 | Sep 22, 2026 | Closed | Claude Code |
| GPT-6 SolNew releaseOpenAIThe original GPT-6 Sol release. GPT-6.1 Sol is the newer version at the same standard input and output rates, with a lower cache-read price.Model and pricingRelease log | $10 output$2 input / 1M tokensChecked October 2, 2026standard; input up to 272K | Sep 22, 2026 | Closed | Codex |
| GPT-6 LunaNew releaseOpenAIThe lower-cost GPT-6 option for routine work and high-volume agent tasks.Model and pricingRelease log | $0.5 output$0.1 input / 1M tokensChecked October 2, 2026standard; input up to 272K | Sep 22, 2026 | Closed | Codex |
| Grok 4.7New releasexAIA 500K-context model with configurable reasoning. Its faster variant is available through Cursor and Grok Build.Release and pricingModel details | $6 output$2 input / 1M tokensChecked October 2, 2026standard; prompts below 200K | Sep 21, 2026 | Closed | Grok Build / Cursor |
| DeepSeek V4.1 FlashNew releaseDeepSeekAn open-weights model with native image input and 1M context. The deepseek-flash API alias serves V4.1; legacy V4 Flash IDs also route here.Release logAPI pricing and aliasesWeights and MIT license | $1.2 output$0.3 input / 1M tokensChecked October 2, 2026peak, uncached; off-peak $0.15 input / $0.60 output; cache hits $0.006 peak / $0.003 off-peak | Sep 10, 2026 | Open | DeepSeek Harness / OpenCode |
| GPT-6 AstraNew releaseOpenAIOpenAI’s most capable broadly available model, combining coding, computer use, and research in long-running tasks.Model and pricingRelease log | $50 output$10 input / 1M tokensChecked October 2, 2026standard; input up to 272K | Sep 3, 2026 | Closed | Codex |
| Muse Spark 1.3New releaseMetaUpdates Spark’s coding and agent behavior, with max reasoning in Standard mode. Available through Meta Model API and Muse Code.The Contributor API variant does not support max reasoning.Spark 1.3 launchAPI pricingModel availability | $4.25 output$1.25 input / 1M tokensChecked October 12, 2026Standard API; cache reads $0.15. Contributor: $0.10 input / $0.20 output, $0.002 cached input; content may train Meta models | Sep 2, 2026 | Closed | Muse Code |
| Gemini 3.8 FlashNew releaseGoogleGoogle’s latest Flash model for software engineering and agent workflows. The separate Cyber variant has restricted access.ReleaseAPI pricing | $3.75 output$0.75 input / 1M tokensChecked October 2, 2026standard promo through Dec 31, 2026; $1.50 input / $7.50 output from Jan 1, 2027 | Sep 2, 2026 | Closed | Gemini CLI / Antigravity |
| Claude Fable 5.1New releaseAnthropicGenerally available for coding and knowledge work. The related Mythos 5.1 model is limited to trusted-access programs.Release and pricingRelease date | $50 output$10 input / 1M tokensChecked October 2, 2026standard API | Sep 1, 2026 | Closed | Claude Code |
| Qwen3.8 Flash-NextNew releaseAlibabaAn open-weight model for repository coding and long-running tool use. Its sparse architecture targets efficient inference over long contexts.Downloadable weights under the Qwen Community license. Qwen3.8 Flash is the hosted version based on these weights, with additional production features.Model card and weightsHosted variant and release date | Self-hostedChecked October 12, 2026Infrastructure or third-party inference costs apply; hosted Qwen3.8 Flash is a related production variant. | Aug 26, 2026 | Open | Qwen Code / compatible agent runtimes |
| GLM 5.3 Flash (Ox Alpha)Z.ai (revealed Aug 26)Released as Ox Alpha before being identified as GLM 5.3 Flash. Volume and availability notes refer to the August snapshot.Description from August snapshotWeights are now downloadable under the MIT license; benchmark and usage figures retain their August snapshot dates.Current API pricingWeights and license | $0.5 output$0.15 input / 1M tokensChecked October 12, 2026Standard API; cache reads $0.03 per million tokens. Faster FlashX service: $0.37 input / $1.25 output. | Aug 20, 2026 | Open | OpenCode (launch partner) |
| Qwen3.8 27BNew releaseAlibabaA dense 27B multimodal model for coding and agent work, with configurable thinking and a smaller deployment footprint than large mixture-of-experts models.Downloadable weights under Apache 2.0; also available through third-party inference providers.Model card and weightsHosted availability and release date | Self-hostedChecked October 12, 2026Infrastructure or third-party inference costs apply; hosted provider rates vary. | Aug 14, 2026 | Open | Qwen Code / local agent runtimes |
| GLM 5.3Z.aiTop-10 coding at a fraction of frontier pricing; 5.2 weights are open, 5.3 Flash weights promised.Description from August snapshotWeights are now downloadable under the GLM-5.3 license; benchmark and usage figures retain their August snapshot dates.Current API pricingWeights and license | $4.4 output$1.4 input / 1M tokensChecked October 12, 2026Standard API; cache reads $0.26 per million tokens | Aug 14, 2026 | Open | ZCode |
| Gemini 3.7 FlashGoogleAugust's fastest riser: near-frontier coding at Flash prices.Description from August snapshot | $1.875 output$0.375 input / 1M tokensSnapshot: August 26, 202675% promo | Aug 13, 2026 | Closed | Gemini CLI / Antigravity |
| Grok 4.6xAICheapest model in the 76+ coding tier; AA flags it for cost efficiency.Description from August snapshot | $6 output$2 input / 1M tokensSnapshot: August 26, 2026 | Aug 12, 2026 | Closed | — |
| DeepSeek V4 ProDeepSeekDeepSeek's reasoning flagship, refreshed mid-August.Description from August snapshotDeepSeek continues to serve V4 Pro after its previously announced September 14 retirement date.Current API pricing and availability | $3.96 output$1.32 input / 1M tokensChecked October 2, 2026peak, uncached; off-peak $0.66 input / $1.98 output | Aug 12, 2026 | Partial | — |
| Muse GlimmerNew releaseMetaA 30B dense text-and-image model distilled from Spark for local agents and coding. Apache 2.0 weights; 128K default context.Glimmer launchArchitecture and license | Self-hostedChecked October 12, 2026No Meta-hosted API rate; infrastructure or third-party hosting costs apply | Aug 10, 2026 | Open | Local agent runtimes |
| Muse Spark 1.2MetaMeta's return to the coding table, ahead of Sonnet 5 on the Coding Index.Description from August snapshotStill available through Meta Model API and Muse Code; Spark 1.3 is the newer version.Model availabilityAPI pricing | $4.25 output$1.25 input / 1M tokensChecked October 12, 2026Standard API; cache reads $0.15. Contributor: $0.10 input / $0.20 output; content may train Meta models | Aug 5, 2026 | Closed | Muse Code |
| Qwen 3.8 MaxAlibabaAlibaba's frontier bid; the 2.4T-parameter open-weights sibling shipped Aug 12.Description from August snapshot | $6 output$2 input / 1M tokensSnapshot: August 26, 2026 | Aug 3, 2026 | Partial | Qwen Code |
| DeepSeek V4 FlashDeepSeekThe volume king of coding: usage spiked ~570% after the July 31 agent-tuned retrain.Description from August snapshotRetired on the DeepSeek API. Its legacy model ID now serves V4.1 Flash at V4.1 prices; the figures below are historical.Retirement and alias routing | $0.1 output$0.05 input / 1M tokensSnapshot: August 26, 2026approx; tiered | Jul 31, 2026 | Partial | — |
| Claude Opus 5Anthropic#1 on AA's Intelligence Index; Claude Code + Opus 5 also tops their Coding Agent Index.Description from August snapshot | $25 output$5 input / 1M tokensSnapshot: August 26, 2026 | Jul 24, 2026 | Closed | Claude Code |
| Kimi K3Moonshot AIThe open-weights frontier for coding, within ~2 points of Opus 5.Description from August snapshot | $15 output$3 input / 1M tokensSnapshot: August 26, 2026 | Jul 16, 2026 | Open | Kimi Code CLI |
| GPT-5.6 SolOpenAIHighest Coding Index score in the August 26 snapshot.Description from August snapshotAPI pricing | $20 output$4 input / 1M tokensChecked October 2, 2026standard promo at least through Nov 21, 2026; input up to 272K | Jul 9, 2026 | Closed | Codex |
| GPT-5.6 TerraOpenAICoding-skewed 5.6 sibling: top-5 coding despite mid-pack general intelligence.Description from August snapshot | $12 output$2 input / 1M tokensSnapshot: August 26, 2026 | Jul 9, 2026 | Closed | Codex |
| GPT-5.6 LunaOpenAIThe volume play: near-Sonnet coding scores at commodity pricing.Description from August snapshot | $1.2 output$0.2 input / 1M tokensSnapshot: August 26, 2026 | Jul 9, 2026 | Closed | Codex |
| Claude Sonnet 5AnthropicThe workhorse mid-tier for agentic coding at a fifth of Fable pricing.Description from August snapshot | $10 output$2 input / 1M tokensSnapshot: August 26, 2026 | Jun 30, 2026 | Closed | Claude Code |
| Claude Fable 5AnthropicThe hard-task tier: #2 on intelligence, priciest mainstream model per token.Description from August snapshot | $50 output$10 input / 1M tokensSnapshot: August 26, 2026 | Jun 9, 2026 | Closed | Claude Code |
| MiniMax M3MiniMaxOpen-weights long-horizon agent model with a 1M context at commodity cost.Description from August snapshotAPI pricing | $1.2 output$0.3 input / 1M tokensChecked October 2, 2026standard up to 512K input; above 512K: $0.60 input / $2.40 output; priority 1.5× | May 31, 2026 | Open | — |
| Composer 2.5CursorCursor’s own coding model. Standard and Fast variants have different token prices.Description from August snapshotCursor pricing and variants | $2.5 output$0.5 input / 1M tokensChecked October 2, 2026Cursor standard; default Fast variant is $3 input / $15 output | May 18, 2026 | Closed | Cursor |
| GPT-5.5OpenAIThe previous OpenAI flagship, still top-10 while the 5.6 family displaces it.Description from August snapshot | $30 output$5 input / 1M tokensSnapshot: August 26, 2026 | Apr 24, 2026 | Closed | Codex |
| Gemini 3.1 ProGoogleGoogle's Pro tier, now behind its own 3.7 Flash on both AA indexes.Description from August snapshot | $12 output$2 input / 1M tokensSnapshot: August 26, 2026 | Feb 19, 2026 | Closed | Gemini CLI / Antigravity |
| Claude Haiku 4.5AnthropicCheap, fast Claude tier for high-volume subagent and edit loops; aging against the 2026 field.Description from August snapshot | $5 output$1 input / 1M tokensSnapshot: August 26, 2026 | Oct 15, 2025 | Closed | Claude Code |
These observations describe July and August 2026. They do not establish current market share or model quality.
High-volume models in the August 26 OpenRouter snapshot. Routed token volume is not a measure of coding quality or total adoption.
| MiMo-V2.5 | Xiaomi | 9.92T/wk | #3 on OpenRouter; Pro variant scores 60.2 on the Coding Index (secondhand) |
| Hy3 | Tencent | 7.16T/wk | open weights; Coding Index 58.8 (secondhand) |
| Nemotron 3 Ultra (free) | NVIDIA | 5.4T/wk | free tier absorbing enormous routed volume |
| GLM 5.2 | Z.ai | 3.22T/wk | open weights (MIT); Coding Index 68.8; superseded by 5.3 but still huge |
| Solar Pro 4 | Upstage | 573B/wk | Aug 12 release; a sleeper climbing the rankings |
Share of model API spend observed among US businesses on Ramp in July 2026. This measures spending within Ramp’s coverage, not the whole market or coding-agent usage.
Emerald = Anthropic, grey = OpenAI. Delta vs the prior month. Source: Ramp AI Index. Limited to spending visible in this source. Missing providers and models should not be read as having zero spend.
The Amplifying Coding Score blends two independent signals: 60% Artificial Analysis Coding Index (Terminal-Bench v2.1 + SciCode, a capability benchmark) and 40% arena coding Elo (OpenLM's Arena+ coding column, Aug 22 snapshot, an LLM-judged preference signal). Each is min-max normalized within this cohort, so 100 is the cohort frontier and 0 the cohort floor, not absolute quality. A model missing one component is scored on the other alone and marked °; missing both means no score and no rank.
Coding Index is a third-party benchmark, quoted at the effort setting noted per row. These are not Amplifying product benchmarks. Scores marked † were recorded from public mirrors of the AA board. They have not been reverified against the current primary board. Releases after the snapshot and models without captured benchmark evidence carry no score. This does not mean that public results are unavailable today.
Tokens per week is OpenRouter's routed volume in the dated snapshot. It covers requests routed through OpenRouter, not the whole market. First-party usage (Claude Code on Anthropic's API, Codex on OpenAI's, Gemini CLI on Google's) never appears there, so frontier-lab volumes are heavily understated. Figures marked ‡ are from a July 22 community snapshot because the model has since left the visible top lists.
Biz spend is the model's share of API spend among US businesses on Ramp's card data (July 2026, per the Ramp AI Index). A dash means no figure was captured, not zero spend. Spending and routed token volume do not feed into the coding score.
Tokenizers differ per provider. Price per token alone does not tell you the cost of a completed task: reasoning, retries, caching, and tool calls all affect the bill. Test models with the same agent, task, and settings before drawing conclusions about cost or quality.
Curated, not exhaustive. Release details and benchmark snapshots are updated separately. Official sources are linked on the new releases; historical benchmark and adoption sources include: Artificial Analysis, OpenRouter rankings, OpenRouter models API, and recent release coverage.
The same model can behave differently with different tools, instructions, and permissions. The Coding Agents Index tracks publicly attributed agent activity: pull requests, commits, merge outcomes, and review patterns.