Developer guide · October 2026One default model for everything is either overpaying or underpowered. Match the task to the model on reasoning depth, token cost and speed, and escalate only when a cheaper model falls short.
OpenAI models
Anthropic models
From frontier horsepower down to cheap, narrow decisions.
OpenAI
Frontier intelligence and high technical complexity
Use when
Long-running, tireless work that needs extreme intelligence and predictable execution with custom skills.
Trade-off
Reserve it for jobs that truly need it. Ultra-fast mode burns plan allowance quickly, and without strong steering it makes expensive mistakes fast.
OpenAI
Heavy data synthesis and high-token workloads
Use when
Big projects with massive context: piles of data, code or documents, or computer use. Persistent without draining weekly limits.
Trade-off
Approaches Astra on some benchmarks at about one-fifth of the token price. Shade task complexity slightly down.
OpenAI · Decisions API
Fast, cheap classification
Use when
Constrained "Type 1" decisions: tagging incoming email, sorting customer records, classifying changes by pricing or release date.
Trade-off
Built for narrow decision structures, not open-ended writing. That focus is what makes it so fast and cheap.
Insight, writing quality and dependable structured execution.
Anthropic
Intuitive synthesis and pattern discovery
Use when
Deep qualitative analysis across messy, massive archives, like months of transcripts and long documents, where people can't see the forest for the trees.
Strength
Finds patterns you didn't know to look for, pulls new insight straight from sources and explains it in clear, well-written prose.
Anthropic
High-quality, natural writing
Use when
Producing a large volume of clean, high-quality written work, and you can give the model time to work.
Strength
Very responsive to steering away from stiff, mannered prose. Slower to generate than Astra.
Anthropic
Efficient, structured execution
Use when
Well-defined, structured tasks with moderate to high complexity.
Strength
Priced the same as Sol. A dependable, highly competent workhorse that reliably gets through the work.
Each model rated 1–5. Cost uses published API prices; speed uses measured throughput where it exists.
| Model | 🧠Brainpower | $CostAPI price, in / out per 1M | 🏃Speed | Best at |
|---|
Context size isn't a differentiator. All six models now take roughly 1 million tokens of input and return up to 128K. Choose on brainpower, cost and speed instead. On OpenAI models, a single request over 272K input tokens is billed at double the input rate.
Prices in US$ per 1M tokens. Context shown as max input / max output.
| Model | Input | Cached input | Output | Context | Throughput | Watch out for |
|---|---|---|---|---|---|---|
| GPT-6 Astra | $10.00 | $1.00 | $50.00 | 1.05M / 128K | Not published | Fast mode: 2× speed at 2× price. Ultrafast: $60 / $300. |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | 1.05M / 128K | Not published | About one-fifth of Astra per task on coding benchmarks. |
| Luna | $0.10 | $0.01 | $0.50 | 1.05M / 128K | Not published | 100× cheaper than Astra. Built for narrow classification. |
| Fable 5.1 | $10.00 | $0.25 | $50.00 | 1M / 128K | ≈ 67 tok/s | Thinking is always on. First token can take minutes at max effort. |
| Opus 5.5 | $4.00 | See pricing page | $20.00 | 1M / 128K | ≈ 106 tok/s | Highest Artificial Analysis score (58). Anthropic’s recommended default. |
| Sonnet 5.5 | $2.00 | See pricing page | $10.00 | 1M / 128K | ≈ 116 tok/s | Uses ~60% more output tokens than Opus at max effort. Measure cost per task. |
OpenAI long-context surcharge: if a single request sends more than 272K input tokens, the whole request is billed at 2× input and 1.5× output. Max input per OpenAI request is about 922K tokens even though the window is 1.05M.
Typical developer tasks, with a prompt to copy and adapt.
AstraSolLunaFable 5.1Opus 5.5Sonnet 5.5Move up a level only when the model below has clearly failed the task.
Luna triage→
Sonnet / Sol build→
Opus reason & review→
Astra / Fable hardest 5%
Put the system prompt, repo map and style rules first and reuse them. Cached input is 10× cheaper on Astra ($1 vs $10) and 40× cheaper on Fable.
Above that the whole request is billed at 2× input. Send the relevant files, not the whole repo.
Per-token price misleads. Sonnet is half Opus’s rate but can use more tokens to finish. Log tokens per ticket and compare.
High effort adds latency and thinking tokens. Save max effort for genuinely hard problems.
On Astra and long agent runs, approve a short plan before it starts editing. A wrong direction at speed is the expensive failure.
Client code, secrets and personal data only go to approved enterprise or API accounts that exclude training. Ask if unsure.