We'll tell you which model to use.
A catalog is not an answer. For each workload we name one default, one cheap alternate and one escape hatch — plus the routing directive that encodes the same decision so you never have to hardcode a model.
| Workload | Default | Cheap alternate | Escape hatch | Directive | Est. $/1M |
|---|---|---|---|---|---|
| Coding agent | anthropic/claude-sonnet-4.5 | alibaba/qwen3-coder-480b | anthropic/claude-opus-4.5 | auto:coding | $3.00 blended |
| Deep reasoning & math | openai/gpt-5.5 | deepseek/deepseek-r1 | anthropic/claude-opus-4.5 | auto:reasoning | $12.50 blended |
| Function calling & agents | openai/gpt-5 | google/gemini-3.1-flash | anthropic/claude-sonnet-4.5 | auto:tools | $1.40 blended |
| Long-context retrieval | google/gemini-3.1-pro | google/gemini-3.1-flash | openai/gpt-5.5 | auto:long-context | $7.00 blended |
| Vision & document parsing | google/gemini-3.1-pro | alibaba/qwen3-vl-235b | openai/gpt-5.5 | auto:vision | $7.00 blended |
| Batch classification & extraction | openai/gpt-5.5-mini | openai/gpt-5.5-nano | anthropic/claude-haiku-4.5 | auto:cost | $0.10–0.40 |
| Realtime & voice turns | groq/llama-4-70b | google/gemini-3.1-flash | openai/gpt-5.5-mini | auto:speed | $0.40 blended |
| High-volume product chat | google/gemini-3.1-flash | deepseek/deepseek-v4 | anthropic/claude-sonnet-4.5 | auto:cost + cache | $0.28–0.55 |
The reasoning behind each pick
Coding agent
auto:coding20–200k context, many tool calls, long diffs, latency tolerant
Highest diff fidelity per dollar; holds multi-file edits without re-reading the repo.
Watch out · Opus is 5x the price for ~3 points of quality — reserve it for failed retries.
→ Full coding agent leaderboardDeep reasoning & math
auto:reasoningShort prompt, long thinking budget, single answer, seconds to minutes
Reasoning tokens dominate the bill, so quality per thinking-token beats headline price.
Watch out · Cap the thinking budget — an uncapped reasoning run can cost 20x a normal call.
→ Full deep reasoning & math leaderboardFunction calling & agents
auto:toolsMany small turns, strict JSON, 5–40 tools in scope
Schema conformance under a wide tool set matters more than raw reasoning score.
Watch out · Provider variance is large here — the same model conforms differently per host.
→ Full function calling & agents leaderboardLong-context retrieval
auto:long-context200k–2M input, small output, recall-critical
Effective recall at depth, not the advertised window, is the deciding number.
Watch out · Above ~400k, chunk-and-rerank is usually cheaper and more accurate than one big call.
→ Full long-context retrieval leaderboardVision & document parsing
auto:visionImages, PDFs, charts and screenshots in, structured data out
Chart and table extraction quality separates models far more than natural-image captioning.
Watch out · Image tokens are billed at very different rates per provider — check the receipt.
→ Full vision & document parsing leaderboardBatch classification & extraction
auto:costMillions of short calls, fixed schema, throughput-bound
At this volume a 3x price difference dwarfs a 2-point quality difference.
Watch out · Cache the system prompt — it is often 80% of the tokens in a batch job.
Realtime & voice turns
auto:speedSub-300ms first token, short turns, interruption-tolerant
TTFT is the whole experience; specialty silicon beats a smarter model that starts late.
Watch out · Specialty hosts have thin capacity — always pin a second upstream for failover.
High-volume product chat
auto:cost + cacheLong-lived sessions, repeated system prompts, cost-sensitive
Semantic caching does more for the bill here than any model swap.
Watch out · Measure cache hit rate before optimising the model — 40% hits beats a cheaper model.