# Model Plan

> Which AI coding model, at which reasoning effort, on which coding plan. Model Plan compares every coding model at its own ★ recommended effort (not just its top setting), shows ability, cost and task time at every effort, and tracks coding plan prices, quotas, off-peak hours and quota resets. Independent, no paid placement; all data is public and read-only.

Data publication 761b476d896fb5f55a1e0d8a, published 2026-10-24T03:36:52.587Z. Numbers change with each publication; read the live API instead of quoting old values.

## Pages
- [AI Coding Model Leaderboard · ranked at each model's recommended effort](https://modelplan.ai/): AI coding model leaderboard: every model compared at its own best-value ★ reasoning effort, not just its top setting. Ability, cost and task time in one view, plus the coding plans that cover them.
- [Recommended reasoning effort for every AI coding model, Low to Max](https://modelplan.ai/effort): Low to Max for every AI coding model: ability, cost and task time at each effort, the ★ recommended one marked, and what one more step up buys and costs.
- [AI coding plan comparison · prices, quotas and covered models](https://modelplan.ai/plans): Coding plans by maker and by model: price, 5-hour and weekly quotas, which models and efforts each plan covers, with sources.
- [Burn Windows · off-peak hours, peaks and quota resets](https://modelplan.ai/windows): Off-peak hours, peak hours, quota resets and promotions for coding plans by week, with calendar feeds.
- [Mission Briefing · latest changes to AI coding models and plans](https://modelplan.ai/updates): Pricing, quota, promotion and reset news for AI coding models and plans, each with its original source.
- [Tool Bay · Model Plan Skill, Claude Code and Codex plugins, CLI and API](https://modelplan.ai/hangar): Bring recommended efforts and plan data into Claude Code, Codex and the terminal: an open-source Skill, plugins, the modelplan CLI, and a key-free read-only API and calendar feeds.
- [Methodology · how ability, cost and the recommended effort are worked out](https://modelplan.ai/methodology): Formulas for ability, reference-task cost, task time and the ★ recommended effort, model selection rules and every data source.
- [About Model Plan](https://modelplan.ai/about): An independent, open project comparing AI coding models and plans: recommended efforts, a leaderboard at those efforts and plan advice, with checkable public data.

## Data (anonymous, read-only JSON)
- [Client digest](https://modelplan.ai/v1/client): current models, every native effort with ability, cost per reference task and time, the recommended effort, per-maker lineups with verdicts, speed tiers, calendar windows and resets
- [Models](https://modelplan.ai/v1/models): full model metrics
- [Coding plans](https://modelplan.ai/v1/plan-groups): plan prices and tiers by product line
- [News](https://modelplan.ai/v1/news?lang=en): pricing, quota and reset news with sources
- [OpenAPI](https://modelplan.ai/openapi.json): every endpoint
- [Calendar](https://modelplan.ai/calendar.ics): off-peak, peak and reset events (ICS)

## Tools
- [Model Plan Skill](https://github.com/modelplan/modelplan/tree/main/skills/model-plan): for Claude Code, Codex and any agent with Skills; answers model and effort questions from the live data and the user's local quota
- [modelplan CLI, Claude Code and Codex plugins](https://github.com/modelplan/modelplan): open source (MIT)
- [Methodology](https://github.com/modelplan/modelplan/blob/main/METHODOLOGY.md)

## Leaderboard at the ★ recommended effort

★ is the best-value effort among a model's own efforts, not its top effort; the top effort is listed separately. Ability is the Model Plan ability score: a frozen scale where 50 is the September 2026 frontier average and 10 points are about one standard deviation; not a percentage, so it can fall below 0 or rise above 100. Cost is the API cost of one reference coding task in USD.

| Model | Maker | Recommended effort | Ability | Cost per task (USD) | Top effort (ability · cost) | Verdict |
| --- | --- | --- | --- | --- | --- | --- |
| [Gemini 4 Argon](https://modelplan.ai/models/gemini-4-argon) | Google | High | 61.6 | $4.26 | same as ★ | Strongest · Best value |
| [Muse Wonder](https://modelplan.ai/models/muse-wonder) | Meta | Max | 58.0 | $7.57 | same as ★ | Strongest · Best value |
| [GPT-6 Astra](https://modelplan.ai/models/gpt-6-astra) | OpenAI | Med | 55.6 | $3.40 | Max · 61.2 · $7.55 | Outclassed |
| [Claude Opus 5.5](https://modelplan.ai/models/claude-opus-5-5) | Anthropic | Med | 55.5 | $2.90 | Max · 67.2 · $11.75 | Strongest · Best value |
| [GPT-6.1 Sol](https://modelplan.ai/models/gpt-6-1-sol) | OpenAI | Med | 55.2 | $0.613 | XHigh · 60.1 · $1.03 | Strongest · Best value |
| [Claude Fable 5.1](https://modelplan.ai/models/claude-fable-5-1) | Anthropic | Med | 54.2 | $5.61 | Max · 62.1 · $13.89 | Outclassed |
| [MiMo-V2.6-Pro](https://modelplan.ai/models/mimo-v2-6-pro) | Xiaomi | Max | 52.6 | $0.379 | same as ★ | Strongest · Best value |
| [Qwen3.8 Max (0902)](https://modelplan.ai/models/qwen3-8-max-0902) | Alibaba | Max | 50.7 | $15.41 | same as ★ | Strongest · Best value |
| [Muse Spark 1.3](https://modelplan.ai/models/muse-spark-1-3) | Meta | Max | 50.5 | $3.25 | same as ★ | Cheapest |
| [Claude Sonnet 5.5](https://modelplan.ai/models/claude-sonnet-5-5) | Anthropic | High | 49.6 | $1.51 | Max · 65.8 · $11.93 |  |
| [Grok 4.7](https://modelplan.ai/models/grok-4-7) | SpaceXAI | High | 48.8 | $5.49 | XHigh · 50.1 · $7.44 | Strongest · Best value |
| [GLM-5.3](https://modelplan.ai/models/glm-5-3) | 智谱 AI | Max | 47.7 | $4.53 | same as ★ | Strongest · Best value |
| [Claude Haiku 5.5](https://modelplan.ai/models/claude-haiku-5-5) | Anthropic | XHigh | 42.7 | $0.487 | Max · 44.3 · $1.46 | Cheapest |
| [GLM-5.3-Flash](https://modelplan.ai/models/glm-5-3-flash) | 智谱 AI | Max | 41.4 | $0.471 | same as ★ | Cheapest |
| [Grok 4.6](https://modelplan.ai/models/grok-4-6) | SpaceXAI | Med | 41.1 | $2.54 | XHigh · 45.3 · $4.17 | Cheapest |
| [Kimi K3](https://modelplan.ai/models/kimi-k3) | Kimi | Max | 39.9 | $5.94 | same as ★ | Cheapest |
| [Gemini 3.8 Flash](https://modelplan.ai/models/gemini-3-8-flash) | Google | High | 39.2 | $2.86 | same as ★ | Cheapest |
| [MiMo-V2.6-Flash](https://modelplan.ai/models/mimo-v2-6-flash) | Xiaomi | Max | 35.1 | $0.178 | same as ★ | Cheapest |
| [GPT-5.6 Sol](https://modelplan.ai/models/gpt-5-6-sol) | OpenAI | Med | 34.5 | $1.48 | Max · 50.3 · $5.99 | Outclassed |
| [GPT-5.6 Terra](https://modelplan.ai/models/gpt-5-6-terra) | OpenAI | High | 29.1 | $0.901 | Max · 45.0 · $3.66 | Outclassed |
| [DeepSeek V4 Flash 0731](https://modelplan.ai/models/deepseek-v4-flash-0731) | DeepSeek | Max | 27.8 | $0.159 | same as ★ | Cheapest |
| [GPT-6 Luna](https://modelplan.ai/models/gpt-6-luna) | OpenAI | High | 26.9 | $0.083 | Max · 37.6 · $0.177 | Cheapest |

## How to read the ★ effort and the ability score

**Is the ★ effort the top effort?** No, by design. ★ is the effort that gives the most for its cost per reference task among a model's own efforts. The top effort's ability and cost are listed beside it on the leaderboard and every model page, and the home field can switch every model to its top effort.

**Why does ★ differ from the vendor's default effort?** A vendor default is a product setting that also weighs responsiveness, rate limits and positioning; taken together, no single ability-for-cost trade-off explains every vendor's defaults. Model Plan shows the vendor default beside ★ as a reference, not as an input.

**Why do cheap models get higher efforts than expensive ones?** Marginal value: one step up costs cents on a cheap model and a dollar or more on an expensive one, so the same ability gain is worth more on the cheap model.

**Why can the ability score be negative or above 100?** It is not a percentage but a frozen scale: 50 is the September 2026 frontier average and 10 points are about one standard deviation. Configurations far below that level (mostly non-reasoning settings) fall below 0, and stronger future models will pass 100. The scale is frozen so scores stay comparable over time.

**What are the known limits?** Ability comes from public benchmarks, both general and coding; cost is one reference task at official API prices (Claude and Codex subscription quotas track API prices too); some times are estimates and are marked. ★ is a product rule calibrated on recorded owner preferences, not a guarantee of task success.

## Coding plans

- [Alibaba Coding Plan](https://modelplan.ai/plans/alibaba-coding-plan)
- [Alibaba Token Plan](https://modelplan.ai/plans/alibaba-token-plan)
- [BigModel Coding Plan](https://modelplan.ai/plans/bigmodel-coding-plan)
- [ChatGPT / Codex](https://modelplan.ai/plans/openai-chatgpt-codex)
- [Claude](https://modelplan.ai/plans/anthropic-claude)
- [Command Code](https://modelplan.ai/plans/commandcode-subscription)
- [Cursor](https://modelplan.ai/plans/cursor-subscription)
- [DeepSeek API](https://modelplan.ai/plans/deepseek-api)
- [Devin](https://modelplan.ai/plans/devin-subscription)
- [GitHub Copilot](https://modelplan.ai/plans/github-copilot)
- [Google Antigravity](https://modelplan.ai/plans/google-antigravity)
- [Kimi / Kimi Code](https://modelplan.ai/plans/kimi-membership)
- [Kiro](https://modelplan.ai/plans/kiro-subscription)
- [MiMo Token Plan](https://modelplan.ai/plans/xiaomi-token-plan)
- [MiniMax M Plan](https://modelplan.ai/plans/minimax-m-plan)
- [Muse Code](https://modelplan.ai/plans/meta-muse-code)
- [Ollama Cloud](https://modelplan.ai/plans/ollama-cloud)
- [OpenCode Go](https://modelplan.ai/plans/opencode-go)
- [OpenCode Zen / Console](https://modelplan.ai/plans/opencode-zen)
- [SuperGrok](https://modelplan.ai/plans/xai-supergrok)
- [The X Subscription](https://modelplan.ai/plans/xai-x-subscription)
- [Volcano Ark Agent Plan](https://modelplan.ai/plans/volcengine-agent-plan)
- [Volcano Ark Coding Plan](https://modelplan.ai/plans/volcengine-coding-plan)
- [Z.ai Coding Plan](https://modelplan.ai/plans/zai-coding-plan)

## Effort ladders

Cost × and time × are relative to the model's ★ recommended effort.

### Gemini 4 Argon (Google)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| ★ High | 61.6 | — | $4.26 | ×1.00 | ×1.00 (est.) |

- Gemini 4 Argon has one measured configuration, so ★ is that one; there is no other effort to compare.

### Muse Wonder (Meta)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| ★ Max | 58.0 | — | $7.57 | ×1.00 | ×1.00 |

- Muse Wonder has one measured configuration, so ★ is that one; there is no other effort to compare.

### GPT-6 Astra (OpenAI)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| Low | 49.1 | -6.5 | $2.21 | ×0.65 | ×0.89 (est.) |
| ★ Med | 55.6 | — | $3.40 | ×1.00 | ×1.00 (est.) |
| High | 58.1 | +2.5 | $3.97 | ×1.17 | ×1.12 (est.) |
| XHigh | 60.6 | +5.0 | $4.75 | ×1.39 | ×1.17 (est.) |
| Max | 61.2 | +5.6 | $7.55 | ×2.22 | ×1.48 (est.) |

- ★ Med is GPT-6 Astra's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.
- One step up to High: +2.5 ability for ×1.17 the cost and ×1.12 the time, not worth it under the calibrated trade-off.
- For maximum quality: Max, ability 61.2 at $7.55 per reference task (×2.22 the cost, ×1.48 the time).

Vendor default (reference only, not an input to ★): Codex Low (suggested start).

### Claude Opus 5.5 (Anthropic)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| Low | 38.9 | -16.6 | $1.23 | ×0.42 | — |
| ★ Med | 55.5 | — | $2.90 | ×1.00 | — |
| High | 60.3 | +4.8 | $3.98 | ×1.37 | — |
| XHigh | 64.0 | +8.6 | $7.17 | ×2.47 | — |
| Max | 67.2 | +11.8 | $11.75 | ×4.04 | — |

- ★ Med is Claude Opus 5.5's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.
- One step up to High: +4.8 ability for ×1.37 the cost, not worth it under the calibrated trade-off.
- For maximum quality: Max, ability 67.2 at $11.75 per reference task (×4.04 the cost).

Vendor default (reference only, not an input to ★): Claude Code Med, same as ★.

### GPT-6.1 Sol (OpenAI)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| Low | 46.3 | -8.8 | $0.414 | ×0.68 | ×0.80 |
| ★ Med | 55.2 | — | $0.613 | ×1.00 | ×1.00 |
| High | 57.8 | +2.6 | $0.852 | ×1.39 | ×1.23 |
| XHigh | 60.1 | +4.9 | $1.03 | ×1.68 | ×1.43 |
| Max | 59.9 | +4.8 | $1.52 | ×2.48 | ×2.25 |

- ★ Med is GPT-6.1 Sol's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.
- One step up to High: +2.6 ability for ×1.39 the cost and ×1.23 the time, not worth it under the calibrated trade-off.
- For maximum quality: XHigh, ability 60.1 at $1.03 per reference task (×1.68 the cost, ×1.43 the time).

Vendor default (reference only, not an input to ★): API Med, same as ★.

### Claude Fable 5.1 (Anthropic)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| Low | 50.2 | -4.0 | $4.71 | ×0.84 | ×0.91 (est.) |
| ★ Med | 54.2 | — | $5.61 | ×1.00 | ×1.00 (est.) |
| High | 58.1 | +3.9 | $7.77 | ×1.39 | ×1.19 (est.) |
| XHigh | 61.4 | +7.2 | $10.66 | ×1.90 | ×1.27 (est.) |
| Max | 62.1 | +7.9 | $13.89 | ×2.48 | ×1.54 (est.) |

- ★ Med is Claude Fable 5.1's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.
- One step up to High: +3.9 ability for ×1.39 the cost and ×1.19 the time, not worth it under the calibrated trade-off.
- For maximum quality: Max, ability 62.1 at $13.89 per reference task (×2.48 the cost, ×1.54 the time).

Vendor default (reference only, not an input to ★): Claude Code High.

### MiMo-V2.6-Pro (Xiaomi)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| ★ Max | 52.6 | — | $0.379 | ×1.00 | — |

- MiMo-V2.6-Pro has one measured configuration, so ★ is that one; there is no other effort to compare.

### Qwen3.8 Max (0902) (Alibaba)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| ★ Max | 50.7 | — | $15.41 | ×1.00 | — |

- Qwen3.8 Max (0902) has one measured configuration, so ★ is that one; there is no other effort to compare.

### Muse Spark 1.3 (Meta)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| XHigh | 44.2 | -6.3 | — | — | ×0.75 (est.) |
| ★ Max | 50.5 | — | $3.25 | ×1.00 | ×1.00 (est.) |

- ★ Max is Muse Spark 1.3's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.

### Claude Sonnet 5.5 (Anthropic)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| Low | 30.9 | -18.7 | $0.547 | ×0.36 | ×0.52 |
| Med | 38.1 | -11.5 | $0.755 | ×0.50 | ×0.69 |
| ★ High | 49.6 | — | $1.51 | ×1.00 | ×1.00 |
| XHigh | 58.8 | +9.2 | $3.69 | ×2.45 | ×2.20 |
| Max | 65.8 | +16.2 | $11.93 | ×7.92 | ×7.13 |

- ★ High is Claude Sonnet 5.5's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.
- One step up to XHigh: +9.2 ability for ×2.45 the cost and ×2.20 the time, not worth it under the calibrated trade-off.
- For maximum quality: Max, ability 65.8 at $11.93 per reference task (×7.92 the cost, ×7.13 the time).

Vendor default (reference only, not an input to ★): Claude Code Med.

### Grok 4.7 (SpaceXAI)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| Low | 38.9 | -9.9 | $2.20 | ×0.40 | ×0.59 (est.) |
| ★ High | 48.8 | — | $5.49 | ×1.00 | ×1.00 (est.) |
| XHigh | 50.1 | +1.3 | $7.44 | ×1.36 | — |

- ★ High is Grok 4.7's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.
- One step up to XHigh: +1.3 ability for ×1.36 the cost, not worth it under the calibrated trade-off.
- For maximum quality: XHigh, ability 50.1 at $7.44 per reference task (×1.36 the cost).

### GLM-5.3 (智谱 AI)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| Low | 27.9 | -19.8 | $1.98 | ×0.44 | ×0.56 (est.) |
| ★ Max | 47.7 | — | $4.53 | ×1.00 | ×1.00 (est.) |

- ★ Max is GLM-5.3's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.

Vendor default (reference only, not an input to ★): API Max, same as ★.

### Claude Haiku 5.5 (Anthropic)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| Low | 23.0 | -19.7 | $0.100 | ×0.21 | ×0.45 |
| Med | 31.3 | -11.4 | $0.179 | ×0.37 | ×0.54 |
| High | 36.5 | -6.2 | $0.299 | ×0.61 | ×0.71 |
| ★ XHigh | 42.7 | — | $0.487 | ×1.00 | ×1.00 |
| Max | 44.3 | +1.6 | $1.46 | ×3.00 | ×3.39 |

- ★ XHigh is Claude Haiku 5.5's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.
- One step up to Max: +1.6 ability for ×3.00 the cost and ×3.39 the time, not worth it under the calibrated trade-off.
- For maximum quality: Max, ability 44.3 at $1.46 per reference task (×3.00 the cost, ×3.39 the time).

Vendor default (reference only, not an input to ★): Claude Code Med.

### GLM-5.3-Flash (智谱 AI)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| ★ Max | 41.4 | — | $0.471 | ×1.00 | ×1.00 (est.) |

- GLM-5.3-Flash has one measured configuration, so ★ is that one; there is no other effort to compare.

### Grok 4.6 (SpaceXAI)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| Low | 28.5 | -12.7 | $1.18 | ×0.47 | ×0.68 (est., single run) |
| ★ Med | 41.1 | — | $2.54 | ×1.00 | ×1.00 (est.) |
| High | 44.9 | +3.7 | $4.12 | ×1.62 | ×1.25 (est., single run) |
| XHigh | 45.3 | +4.1 | $4.17 | ×1.64 | ×1.63 (est., single run) |

- ★ Med is Grok 4.6's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.
- One step up to High: +3.7 ability for ×1.62 the cost and ×1.25 the time (single run; AA generation time ×1.31), not worth it under the calibrated trade-off.
- For maximum quality: XHigh, ability 45.3 at $4.17 per reference task (×1.64 the cost, ×1.63 the time (single run)).

### Kimi K3 (Kimi)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| Low | 11.7 | -28.2 | $3.27 | ×0.55 | — |
| ★ Max | 39.9 | — | $5.94 | ×1.00 | — |

- ★ Max is Kimi K3's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.

### Gemini 3.8 Flash (Google)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| Low | 23.5 | -15.7 | — | — | — |
| Med | 36.7 | -2.5 | $2.79 | ×0.98 | ×0.80 (est.) |
| ★ High | 39.2 | — | $2.86 | ×1.00 | ×1.00 |

- ★ High is Gemini 3.8 Flash's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.

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

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| ★ Max | 35.1 | — | $0.178 | ×1.00 | — |

- MiMo-V2.6-Flash has one measured configuration, so ★ is that one; there is no other effort to compare.

### GPT-5.6 Sol (OpenAI)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| None | 12.4 | -22.1 | — | — | — |
| Low | 23.3 | -11.2 | $0.698 | ×0.47 | ×0.52 (est., single run) |
| ★ Med | 34.5 | — | $1.48 | ×1.00 | ×1.00 (est.) |
| High | 40.9 | +6.4 | $2.47 | ×1.66 | ×1.40 (est., single run) |
| XHigh | 44.3 | +9.8 | $3.72 | ×2.51 | ×2.65 (est., single run) |
| Max | 50.3 | +15.8 | $5.99 | ×4.04 | ×5.32 (est., single run) |

- ★ Med is GPT-5.6 Sol's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.
- One step up to High: +6.4 ability for ×1.66 the cost and ×1.40 the time (single run; AA generation time ×1.72), not worth it under the calibrated trade-off.
- For maximum quality: Max, ability 50.3 at $5.99 per reference task (×4.04 the cost, ×5.32 the time (single run)).

Vendor default (reference only, not an input to ★): API Med, same as ★.

### GPT-5.6 Terra (OpenAI)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| None | 2.9 | -26.2 | $0.398 | ×0.44 | — |
| Low | 16.9 | -12.2 | $0.399 | ×0.44 | ×0.56 (est., single run) |
| Med | 21.9 | -7.3 | $0.512 | ×0.57 | ×0.72 (est., single run) |
| ★ High | 29.1 | — | $0.901 | ×1.00 | ×1.00 (est.) |
| XHigh | 35.9 | +6.7 | $1.56 | ×1.73 | ×2.01 (est., single run) |
| Max | 45.0 | +15.9 | $3.66 | ×4.06 | ×5.68 (est., single run) |

- ★ High is GPT-5.6 Terra's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.
- One step up to XHigh: +6.7 ability for ×1.73 the cost and ×2.01 the time (single run; AA generation time ×1.72), not worth it under the calibrated trade-off.
- For maximum quality: Max, ability 45.0 at $3.66 per reference task (×4.06 the cost, ×5.68 the time (single run)).

Vendor default (reference only, not an input to ★): API Med.

### DeepSeek V4 Flash 0731 (DeepSeek)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| ★ Max | 27.8 | — | $0.159 | ×1.00 | ×1.00 |

- DeepSeek V4 Flash 0731 has one measured configuration, so ★ is that one; there is no other effort to compare.

### GPT-6 Luna (OpenAI)

| Effort | Ability | vs ★ | Cost per task | Cost × | Time × |
| --- | --- | --- | --- | --- | --- |
| None | -2.0 | -28.9 | $0.032 | ×0.39 | — |
| Low | 4.3 | -22.6 | $0.013 | ×0.16 | ×0.79 (est., single run) |
| Med | 20.8 | -6.1 | $0.050 | ×0.60 | ×0.35 (est., single run) |
| ★ High | 26.9 | — | $0.083 | ×1.00 | ×1.00 (est.) |
| XHigh | 30.5 | +3.6 | $0.120 | ×1.46 | ×2.34 (est., single run) |
| Max | 37.6 | +10.7 | $0.177 | ×2.15 | ×4.07 (est., single run) |

- ★ High is GPT-6 Luna's best-value effort: ability weighed against cost per reference task and time within this model; it is not the top effort.
- One step up to XHigh: +3.6 ability for ×1.46 the cost and ×2.34 the time (single run; AA generation time ×1.40), not worth it under the calibrated trade-off.
- For maximum quality: Max, ability 37.6 at $0.177 per reference task (×2.15 the cost, ×4.07 the time (single run)).

Vendor default (reference only, not an input to ★): Codex High (suggested start), same as ★; API Med.

## Recent news
- 2026-10-23 Kimi K3.1 Pro begins rolling out on Kimi AI web and mobile (https://x.com/testingcatalog/status/2113663517056504159)
- 2026-10-24 Moonshot releases Kimi K3.1 Pro: 3T-class open-weights model stronger at coding, 3D and finance, weights to open before November 4 (https://x.com/shao__meng/status/2113799036004450807)
- 2026-10-20 Codex responds to abnormal quota usage: limit mismatch for about 40,000 accounts fixed with compensation (https://x.com/thsottiaux/status/2112592264686149749)
- 2026-09-29 OpenAI launches $500/month Pro plan with exclusive Astra Ultrafast fastest model (https://www.ithome.com/1/008/525.htm)
- 2026-09-29 Reported: OpenAI to reopen $200 Pro plan tomorrow with usage-based calculation (https://x.com/testingcatalog/status/2104890394936852676)
- 2026-09-29 OpenAI Reopens $200 Pro Plan but Halves API Credits (https://the-decoder.com/openai-reopens-its-200-pro-plan-but-cuts-api-credits-in-half-as-it-nudges-users-toward-pay-per-use/)
- 2026-09-29 OpenAI to reopen $200 ChatGPT Pro subscription tomorrow with adjusted usage calculation (https://www.ithome.com/1/008/295.htm)
- 2026-09-29 Codex Pro $200 Subscription to Reopen, Usage Recalculated as Half of API-Equivalent Spend (https://x.com/thsottiaux/status/2104823812042940713)
- 2026-10-22 Auto mode to become default on Claude web, desktop and mobile (https://x.com/claudeai/status/2113360845409583307)
- 2026-10-21 OpenAI GPT 6.1 Models Coming to Google Cloud and Antigravity (https://arstechnica.com/ai/2026/10/openais-latest-ai-models-are-coming-to-google-cloud-and-antigravity/)
- 2026-10-07 Anthropic adds monthly Platform API credits to Claude Max and Team plans (https://x.com/ClaudeDevs/status/2107895957933408429)
- 2026-09-29 ChatGPT Pro adds $500/month Pro 500 tier, Pro 200 reopens to new subscribers (https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers)
- 2026-09-29 OpenAI Codex Lead Previews Pro $200 Plan Reopening Tomorrow with Usage Calculation Changes (https://x.com/thsottiaux/status/2104838506363408740)
- 2026-10-19 OpenCode Console breached, leaking API keys that ran up unauthorized charges (https://techcrunch.com/2026/10/19/hackers-hit-ai-coding-tool-opencode-running-up-charges-on-users-ai-accounts/)
- 2026-10-23 Kimi K3.1 Pro arrives in OpenCode Go (https://x.com/opencode/status/2113699251616412056)
- 2026-10-23 Moonshot releases Kimi K3.1 Pro: 3T-scale open-weights model with 1M-token context (https://www.kimi.com/blog/kimi-k3-1-pro)
- 2026-10-22 Claude models officially available on Google Cloud (https://x.com/claudeai/status/2113380122468454765)
- 2026-10-21 Claude Code free credit reset expires tomorrow (https://x.com/lydiahallie/status/2113007239476977744)
- 2026-10-15 Gemini 4 Argon rolls out to Google AI Pro subscribers in Antigravity (https://x.com/testingcatalog/status/2110872881261601280)
- 2026-09-29 Devin now supports login with ChatGPT Plus or Pro subscriptions to cover OpenAI model usage (https://x.com/cognition/status/2104996240190792029)