OpenCodeGoEvals/MODELS.md

# OpenCode Go — what each model is for

Short, practical notes on the 23 models in the OpenCode Go catalog: what each one is targeted for (agentic work, coding, data analysis, image understanding, deep research, high-volume chores), what it costs, and what to watch out for.

Evaluated scores come from Artificial Analysis (single consistent harness) plus vendor model cards, retrieved 2026-10-21. The interactive version with the full task matrix lives at `site/index.html`.

> No model in the Go catalog generates images. Every Go model is text-output; several accept image, video, PDF or audio as **input**.

## Quick picks

| If you need… | Use |
|---|---|
| The strongest model overall, for hard agentic + coding work | Gemini 4 Argon Contributor |
| The best value per token for everyday coding agents | DeepSeek V4.1 Flash |
| Zero-cost exploration (limited time) | MiniMax M3.1 Flash Free, Beam Free, LongCat 2.5 Preview Free |
| Long-horizon agent runs with multimodal input | Muse Spark 1.3 Contributor, MiMo-V2.6-Pro |
| Repo-scale code changes | Kimi K2.7 Code, GLM-5.3, Qwen3.8 Max |
| Spreadsheets / data analysis deliverables | Kimi K3 (SpreadsheetBench 34.8), Qwen3.8 Max (GDPval 58.2) |
| Reading screenshots, charts and documents | Qwen3.8 Flash, GLM-5.3-Flash, MiniMax M3 |
| High-volume, latency-sensitive work | Claude Haiku 5.5, GPT-6 Luna, MiMo-V2.5 |

## The catalog

### MiniMax M3.1 Flash Free

**MiniMax** · released 2026-10-19 · $0 in / $0 out per 1M tokens · **free tier (limited time)** · 1M context

I/O: text, image, video in → text out

Targeted for: **Coding, Multimodal, Agentic, Free tier**

Frontier multimodal coding model with a 1M-token context window and tunable thinking depth. The newest model to land in Go (released 19 Oct 2026) and the fastest-growing free option in the catalog.

Day-to-day coding-agent work on real repositories: reading long codebases, editing files, running tools. Image and video input make it usable for UI/screenshot-driven work (reading a rendered page, describing an asset). Because thinking depth is tunable, it can be dialed down for cheap, high-volume turns and up for hard problems.

*Evaluated:* Intelligence Index 42.7 · FrontierCode 38.6 · TB4.0 25.8 · AutomationBench 72.5 · HLE 48.0

**Caveats**

- Free for a limited time, then billed like the rest of the catalog.
- Text output only — it analyses images/video but does not generate them.

### Beam Free

**Reflection AI** · released 2026-10-05 · $0 in / $0 out per 1M tokens · **free tier (limited time)** · 1M context

I/O: text only

Targeted for: **Coding, Reasoning, Long-horizon agents, Free tier**

Reflection's first open-weight model: a 501B-parameter sparse Mixture-of-Experts with 23B active per token, built for coding, reasoning and long-horizon agentic workloads.

Cost-efficient coding and agent loops. The MoE design activates only 23B parameters per token, so it is a cheap way to run a lot of terminal work, tool calls and multi-step reasoning. Text-only, so it is a pure code/reasoning worker rather than a perception model.

*Evaluated:* Intelligence Index 30.3 · SciCode 48.5 · TB4.0 3.0 · AutomationBench 48.5 · HLE 34.7

**Caveats**

- Free for a limited time and unavailable in a long list of countries (see the OpenCode Go docs for the exclusion list).
- Weights are slated for an Apache 2.0 release, with the technical report and model card following.

### LongCat 2.5 Preview Free

**Meituan** · released 2026-09-25 · $0 in / $0 out per 1M tokens · **free tier (limited time)** · 1M context

I/O: text, image in → text out

Targeted for: **Coding, Agentic, Image understanding, Free tier**

Meituan's multimodal reasoning model for coding and agent tasks, with image understanding and a 1M-token context window.

Agents that have to read a picture to act: UI automation, screenshot-to-code, chart and document understanding, then long runs of tool use against a 1M-token window. Pairs well with a free-tier workflow — long exploratory sessions without burning a paid model's allowance.

*Evaluated:* no public third-party evaluation published yet.

**Caveats**

- Free for a limited time.
- Preview release: scores and behaviour may shift.
- Text output only — no image generation.

### Gemini 4 Argon Contributor

**Google** · released 2026-10-13 · $4 in / $20 out per 1M tokens · 1M context

I/O: text, image, video, audio, PDF in → text out

Targeted for: **Frontier coding, Deep research, Legal & financial work, Security, Multimodal**

Google's flagship reasoning model for complex software engineering, deep math and science, and long-horizon autonomous agents. The highest overall score in the Go catalog.

Hardest frontier work: large codebase refactors and optimisation (Google used it to optimise a 200,000-line Python codebase overnight), financial research, legal drafting, deep math/science, and autonomous security work such as vulnerability patching. It is also the model to reach for when a task needs to stay coherent across a very long autonomous run — the 1M window and long max-output budgets are the point, not the per-token price.

*Evaluated:* Intelligence Index 52.6 · FrontierCode 43.5 · SciCode 61.8 · TB4.0 57.1 · AutomationBench 77.5 · HLE 57.1

**Caveats**

- Contributor tier: your prompts and completions may be used to train future Google models. Not zero-data-retention.
- Most expensive model in Go at $4 / $20 per 1M tokens.
- Currently priced at a promotional 2× allowance for a limited time.

### Grok 4.7

**xAI** · released 2026-09-21 · $2 in / $6 out per 1M tokens · 500K context

I/O: text, image, PDF in → text out

Targeted for: **Long-running agents, Coding, Knowledge work, Visual projects**

xAI's frontier model for long-running agents, coding, knowledge work, and visual projects.

Sustained agent sessions that mix code and document/visual inputs — research that ends in a written deliverable, codebase work with images in the loop, or an analysis that has to survive a long run without losing the thread. Priced at the top of the open-model tier but with a 500K window that suits multi-file, multi-hour tasks.

*Evaluated:* Intelligence Index 46.4 · FrontierCode 33.8 · TB4.0 25.8 · AutomationBench 65.6 · GDPval 60.6

**Caveats**

- Premium price ($2 / $6) with a lower $15 monthly allowance in the Go plan.
- Price doubles above 200K tokens of context.

### GPT-6 Luna

**OpenAI** · released 2026-09-22 · $0.1 in / $0.5 out per 1M tokens · 1.05M context

I/O: text, image, PDF in → text out

Targeted for: **High-volume coding, Agents, Document work, Cheap**

OpenAI's most efficient model for focused, high-volume tasks, at the price of a small open model.

The default workhorse when you want GPT-class behaviour without GPT-class spend: everyday coding-agent turns, small refactors, summarising long documents (1.05M window), classifying and routing. In Go it is one of the best price/performance points, so it is a good default for background agents and a fallback when a heavier model is unnecessary.

*Evaluated:* Intelligence Index 38.1 · FrontierCode 41.9 · AutomationBench 53.2 · LCR 83.3 · MMMU-Pro 79.7

**Caveats**

- Price steps up above 272K tokens of context.
- Not a reasoning-tier model — deep analysis work should go to a stronger model.

### Claude Haiku 5.5

**Anthropic** · released 2026-10-07 · $0.1 in / $0.5 out per 1M tokens · 1M context

I/O: text, image, PDF in → text out

Targeted for: **Fast agents, Classification, Responsive UX, Cheap**

Fast Claude model for responsive assistance, classification, and lightweight agents.

Anything that has to answer fast and often: interactive completions, classification and extraction, lightweight sub-agents, and UI-style assistance where latency matters more than depth. Anthropic's tool-calling behaviour is the draw for lightweight agents.

*Evaluated:* Intelligence Index 43.4 · FrontierCode 41.8 · TB4.0 32.8 · Analyst Agent 53.8 · GDPval 55.9

**Caveats**

- Price steps up above 100K tokens of context (5×).
- Requests are retained for 30 days under Anthropic's data policy — not zero-data-retention.

### Muse Spark 1.3 Contributor

**Meta** · released 2026-09-02 · $0.1 in / $0.2 out per 1M tokens · 1M context

I/O: text, image, video, PDF, audio in → text out

Targeted for: **Long-horizon agents, Multi-agent workflows, Coding, Multimodal, Cheap**

Meta's multimodal reasoning model for long-running agentic, multi-agent and coding workflows, improved over 1.2 for long-horizon agent collaboration, instruction following and coding efficiency.

Multi-agent and long-horizon pipelines: several specialised agents handing work to each other over a 1M window, with the widest input surface in Go (image, video, PDF and audio all accepted). Cheap enough ($0.10 / $0.20) to keep several agents running at once — it is the pick when the architecture, not the single call, is doing the work.

*Evaluated:* Intelligence Index 48.1 · TB2.1 84.3 · AutomationBench 57.9 · GPQA 93.5 · GDPval 59.0

**Caveats**

- Contributor tier: your prompts and completions may be used to train future Meta models, and availability is limited to regions permitted by Meta's Geographic Use Policy.
- Text output only.

### Step 5 Preview

**StepFun** · released 2026-09-16 · $1 in / $2.7 out per 1M tokens · 1M context

I/O: text, image, video in → text out

Targeted for: **Coding, Professional knowledge work, Multimodal, Open weights**

StepFun's next-generation flagship base model for coding and professional knowledge work, with native text, image and video input and a 1M-token context window.

Coding plus deliverables: a base model strong enough to write production code and also produce the spreadsheets, reports and documents that surround it. Native image and video input support it works from screenshots and clips directly.

*Evaluated:* Intelligence Index 43.7 · FrontierCode 38.4 · TB4.0 33.3 · AutomationBench 51.0 · LCR 88.3

**Caveats**

- Preview release — expect the numbers to move.
- Currently priced with a promotional 4× allowance for a limited time.

### DeepSeek V4.1 Flash

**DeepSeek** · released 2026-09-10 · $0.15 in / $0.6 out per 1M tokens · 1M context

I/O: text, image in → text out

Targeted for: **Agentic coding, Reasoning, Cheap, High volume**

DeepSeek's V4.1 Flash: built for reasoning and agentic coding, and the single most-used model in OpenCode by token volume.

The default engine for long coding-agent runs. DeepSeek's own card lists Terminal-Bench, DeepSWE, CyberGym and AutomationBench scores, i.e. it is tuned for terminal work, repo-level fixes and security tasks rather than chat. With cache reads at ~0.3¢/1M it is also the cheapest way to move large volumes of agent traffic.

*Evaluated:* Intelligence Index 39.5 · FrontierCode 44.0 · AutomationBench 68.9 · TB4.0 26.8 · LCR 84.0

**Caveats**

- Text output only.
- Price doubles during DeepSeek's peak hours (01:00–04:00 and 06:00–10:00 UTC, Mon–Fri).

### DeepSeek V4 Pro

**DeepSeek** · released 2026-04-24 · $0.66 in / $1.98 out per 1M tokens · 1M context

I/O: text only

Targeted for: **Long agent runs, Coding, Math & science, Open weights**

Open MiE flagship with million-token context for coding and long agent runs.

The heavier DeepSeek: multi-hour agent sessions and research runs that need a large window, plus the maths and science scores (IMOAnswerBench, HMMT, LiveCodeBench) for technical work. Text-only, so it is a reasoning/code model rather than a perception model.

*Evaluated:* Intelligence Index 36 · FrontierCode 38.8 · TB2.1 78.7 · GPQA 92.8 · AutomationBench 56.7

**Caveats**

- Text input only.
- Price doubles during DeepSeek's peak hours.

### Hy4 preview

**Tencent** · released 2026-08-28 · $0.834 in / $2.501 out per 1M tokens · 1M context

I/O: text only

Targeted for: **Agentic workflows, Productivity, Complex task execution, Open weights**

A next-generation productivity model with significantly enhanced agent and complex task execution capabilities.

Productivity and workflow automation: multi-step office/enterprise tasks, tool orchestration and long-running assistants. Tencent positions it around agentic capability and complex task execution rather than raw chat quality, with a 1M window for large corpora.

*Evaluated:* no public third-party evaluation published yet.

**Caveats**

- Preview release with no third-party evaluations published yet, so performance is unverified outside the vendor's own numbers.
- Text input only.

### GLM-5.3

**Zhipu** · released 2026-08-14 · $1.4 in / $4.4 out per 1M tokens · 1M context

I/O: text only

Targeted for: **Long-horizon coding, Agents, Project delivery, Cyber/security**

Flagship GLM model for long-horizon coding, agents, and complex project delivery.

Owning a project rather than a ticket: sustained multi-file work, terminal and repo tasks, and the security-agent track (CyberGym, ExploitGym, ExploitBench) where GLM scores are among the strongest in Go. Text-only, and one of the pricier models, so it is aimed at the tasks that genuinely need a flagship.

*Evaluated:* Intelligence Index 44.8 · FrontierCode 40.1 · TB2.1 83.9 · AutomationBench 62.2 · GPQA 91.7

**Caveats**

- Text input only.
- Expensive per token ($1.40 / $4.40) with a $15 monthly allowance in the Go plan.

### GLM-5.3-Flash

**Zhipu** · released 2026-08-26 · $0.15 in / $0.5 out per 1M tokens · 1M context

I/O: text, image, video, PDF in → text out

Targeted for: **Efficient coding, Long-horizon agents, Multimodal, Cheap**

Native multimodal GLM model for efficient coding and long-horizon agent tasks.

The cheap twin of the flagship: same family behaviour, multimodal input, and one of the better Terminal-Bench scores in Go per dollar. It is the sensible pick for long agent loops that need to see images, documents or video but do not need frontier depth.

*Evaluated:* Intelligence Index 41.8 · FrontierCode 41.0 · TB2.1 84.3 · AutomationBench 60.4 · GPQA 91.2

**Caveats**

- Text output only.
- Loses a little strength on the hardest reasoning tasks compared with GLM-5.3.

### Qwen3.8 Max

**Alibaba** · released 2026-08-03 · $2 in / $6 out per 1M tokens · 1M context

I/O: text, image, video, PDF in → text out

Targeted for: **Coding, Professional work, Multimodal, Long-horizon agents**

A 2.4-trillion-parameter MoE flagship for coding, professional work, multimodal understanding, and long-horizon agentic workflows.

Broad, heavy-duty work where a single model has to cover the whole job: writing code, producing professional deliverables (the GDPval/JobBench scores are strong), and reasoning over images, video and PDFs in the same session. Also the strongest general agent in Go on the terminal and tool-use benchmarks.

*Evaluated:* Intelligence Index 45.4 · FrontierCode 43.2 · TB2.1 88.8 · GPQA 92.8 · MMMU-Pro 82.8

**Caveats**

- Priced at $2 / $6 with a $15 monthly allowance.
- Cache writes are charged ($2.50 / 1M), so long sessions with big prompt churn cost more than the headline rate.

### Qwen3.8 Flash

**Alibaba** · released 2026-08-26 · $0.15 in / $0.47 out per 1M tokens · 1M context

I/O: text, image, video in → text out

Targeted for: **Visual reasoning, Documents, Agent tasks, Cheap**

Qwen vision-language model for visual reasoning, documents, and agent tasks.

Perception-heavy work at a low price: reading screenshots and diagrams, extracting from documents, and driving agents that need to see what is on screen. It pairs a 1M window with sub-20¢ input pricing, which makes it a good bulk worker for visual pipelines.

*Evaluated:* Intelligence Index 39.8 · TB2.1 86.1 · GPQA 92.3 · MMMU-Pro 79.8 · GDPval 56.4

**Caveats**

- Text output only — no image generation.
- Evaluated publicly as 'Qwen3.8-Flash-Next' in third-party benchmark tables.

### Qwen3.7 Plus

**Alibaba** · released 2026-06-02 · $0.4 in / $1.6 out per 1M tokens · 1M context

I/O: text, image, video in → text out

Targeted for: **Long-context agents, Visual inputs, Coding, Multimodal**

Multimodal Qwen workhorse for long-context agents, visual inputs, and coding.

A steady middle tier: long-context agent sessions, image/video understanding and ordinary coding, without the flagship price. Instruction-following scores are its standout metric, which matters for agents driven by precise specs.

*Evaluated:* Intelligence Index 25.2 · TB2.1 61.0 · GPQA 90.0 · IFBench 78.0 · τ²-bench 93.0

**Caveats**

- Text output only.
- Public evaluations are thinner than for the Qwen3.8 line, so treat third-party comparisons with care.

### Kimi K3

**Moonshot** · released 2026-07-16 · $3 in / $15 out per 1M tokens · 1M context

I/O: text, image, video in → text out

Targeted for: **Long-horizon agents, Coding, Web research, Spreadsheets, Open weights**

Multimodal Kimi model with a 1M context and toggleable max-effort thinking for long-horizon agent work.

The Kimi tier built for endurance: long-horizon agent runs, web research (BrowseComp), coding (FrontierSWE, DeepSWE, Terminal-Bench), and office deliverables — it scores well on SpreadsheetBench, which makes it the pick for data-analysis and spreadsheet work in Go. Thinking effort can be turned up for hard problems or down for speed.

*Evaluated:* Intelligence Index 43.6 · FrontierCode 43.6 · TB2.1 85.0 · GPQA 93.5 · BrowseComp 91.2

**Caveats**

- Most expensive model in Go at $3 / $15 per 1M tokens.
- Terminal-Bench 4.0 is much weaker than its Terminal-Bench 2.1 result — hard-shell tasks still cost it a lot.
- Text output only.

### Kimi K2.7 Code

**Moonshot** · released 2026-06-12 · $0.95 in / $4 out per 1M tokens · 256K context

I/O: text, image, video in → text out

Targeted for: **Repo-scale coding, Long-horizon repo work, Tool use**

Coding-focused Kimi model, stronger on long-horizon repo work with less overthinking.

Purpose-built for code: sustained repository work, refactors and multi-file changes where a general model wastes tokens deliberating. The tool-use (τ-bench) score is very high, so it is also good at driving MCP servers and CLI tools. Use it when the deliverable is a diff, not an essay.

*Evaluated:* Intelligence Index 25.8 · TB2.1 67.4 · τ²-bench 90.1 · GPQA 89.6

**Caveats**

- 256K context — roughly a quarter of the 1M models.
- Narrower general-purpose ability than the flagship Kimi.

### MiMo-V2.6-Pro

**Xiaomi** · released 2026-09-22 · $0.435 in / $0.87 out per 1M tokens · 1M context

I/O: text, image, audio, video in → text out

Targeted for: **Coding agents, Multimodal, Automation, Long context**

Stronger MiMo Pro tier for multimodal reasoning and coding-agent execution.

Coding agents that also have to look at things: screenshots, video, audio alongside code. Xiaomi's own card highlights strong AutomationBench and terminal scores, so it is aimed at long mechanical agent loops as much as at authoring code.

*Evaluated:* Intelligence Index 46.3 · FrontierCode 45.4 · TB4.0 34.8 · AutomationBench 58.6 · LCR 86.3

**Caveats**

- Text output only.
- The Free tier of this family (MiMo-V2.6-Flash) sits outside the Go plan — check the default provider list.

### MiMo-V2.5-Pro

**Xiaomi** · released 2026-04-22 · $0.435 in / $0.87 out per 1M tokens · 1M context

I/O: text, image, audio, video in → text out

Targeted for: **Multimodal coding agents, Long-context automation**

Previous-generation MiMo Pro tier for multimodal reasoning and coding-agent execution.

Now the value option in the MiMo line — same interface as 2.6 Pro with materially weaker agentic and cyber benchmarks. Reasonable when you need multimodal input cheaply and the task is not frontier-hard.

*Evaluated:* Intelligence Index 26 · TB2.1 65.2 · τ²-bench 94.2 · GPQA 86.6

**Caveats**

- Superseded by MiMo-V2.6-Pro; scores are materially lower on agentic and security tasks.
- Text output only.

### MiMo-V2.5

**Xiaomi** · released 2026-04-22 · $0.14 in / $0.28 out per 1M tokens · 1M context

I/O: text, image, audio, video in → text out

Targeted for: **Budget multimodal agents, Long context**

Open MiMo model for multimodal coding agents and long-context automation.

The cheapest way to get multimodal, long-context agent behaviour in Go: 14¢/28¢ per 1M. Best suited to bulk automation — many small steps over a large context — rather than to hard single-shot questions.

*Evaluated:* Intelligence Index 21.8 · TB2.1 63.7 · τ²-bench 90.6 · GPQA 84.9

**Caveats**

- Lowest tier of the MiMo line; weak on the hardest agentic benchmarks.
- Text output only.

### MiniMax M3

**MiniMax** · released 2026-06-01 · $0.3 in / $1.2 out per 1M tokens · 1M context

I/O: text, image, video in → text out

Targeted for: **Long-context coding, Perception, Agent planning, SWE tasks**

MiniMax multimodal model for long-context coding, perception, and agent planning.

Repo-scale SWE work with perception in the loop: MiniMax reports one of the strongest SWE-Bench Verified results in Go, plus good browser and desktop-computer scores (BrowseComp, MCP Atlas, OSWorld). Use it when the agent has to look at a screen and then change code.

*Evaluated:* Intelligence Index 29.2 · SWE-Bench Verified 80.5 · Terminal-Bench 66.0 · BrowseComp 83.5 · OSWorld 70.1

**Caveats**

- Text output only.
- Weaker than MiniMax M3.1 Flash on frontier reasoning and long-horizon tasks.

## Free models outside the Go plan

### MiMo-V2.6-Flash Free (Xiaomi)

*Where:* Default OpenCode provider (not in the Go plan)

Multimodal coding agents and long-context automation at no token cost — the free counterpart to MiMo-V2.6-Pro. Text, image, audio and video in; text out.

_Unlimited-style free tier; use it for exploration and bulk mechanical work._

### Nemotron 3.5 Lightning Free (NVIDIA)

*Where:* Default OpenCode provider

Fast, reliable agentic tasks across enterprise workloads — an MoE tuned for dependable tool use rather than flash reasoning, at 256K context.

_Text only; good fit for internal automation and routing._

### Big Pickle (Unclear / community)

*Where:* Default OpenCode provider

An inexpensive general-purpose slot with no vendor model card published. Treat as an experiment, not as a recommended production model.

_No published benchmark card — verify behaviour before depending on it._

## Sources

- [OpenCode Go docs](https://opencode.ai/docs/go) — pricing, allowances, privacy and retention per model
- [OpenCode Data](https://opencode.ai/data) — catalog metadata, context windows, usage
- [Artificial Analysis](https://artificialanalysis.ai) — benchmark scores and Intelligence Index
- Vendor announcements: Google (Gemini 4 Argon), Reflection AI (Beam), Meituan (LongCat), MiniMax (M3.1 Flash), xAI (Grok 4.7), StepFun (Step 5)