[octo-patch]

Reason: Make MiniMax-M3.1-Flash the default for new configurations while retaining MiniMax-M3 and preserving explicit saved model IDs.

Summary

MiniMax setup and ordinary built-in model choices now start with Flash, followed by M3. Legacy model metadata and aliases remain available for saved configurations and explicit lookup, and custom models remain selectable.

Changes

• Register Flash with a 1,000,000-token context, 524,288-token maximum output, and $0.15/$0.60 per-million input/output prices; update all MiniMax setup defaults.
• Separate legacy built-in MiniMax entries from ordinary catalog lists without rewriting saved configuration or model IDs.
• Correct M3's standard short-context catalog prices and document the limits of flat-rate cost estimates.
• Preserve unsigned thinking history for Flash and M3 on the Messages transport, including streaming requests.
• Preserve user text/image blocks in streaming OpenAI-compatible requests using the same conversion as nonstreaming requests.
• Add offline HTTP request-capture tests through the driver factory for both documented regional URL paths, both transports, stream/nonstream, defaults, explicit M3, saved legacy/custom IDs, images, tools, and reasoning history.

Testing

• cargo fmt --all -- --check
• git diff --check
• cargo clippy --workspace -- -D warnings passes (12 double_must_use errors in openfang-types)
• cargo test --workspace passes on Ubuntu, macOS, and Windows
• cargo audit passes (30 vulnerabilities reported)
• Live integration tested (not run; no paid inference)

Validation is for head c7f5729e72efd77f463b2ee7ae1f6473464ef841 in CI run 39562323370. The platform workspace test commands pass after correcting the new fixture's DriverConfig initialization and preserving user image blocks in streaming OpenAI-compatible requests. All three compile/check jobs, formatting, secrets scan, and installer smoke pass. Clippy and Security Audit remain unresolved; no unchanged-base reproduction was performed, so these failures are not classified as baseline.

No local builds or installs were run because heavy setup remains paused below the 110 GiB free-disk threshold. The existing private formatter and git diff --check passed. The minimax_wire tests use offline HTTP fixtures; they do not validate regional service availability or live model behavior.

Thinking and effort controls remain omitted by these drivers for MiniMax, using the service defaults. This change does not introduce an effort selector, native Responses transport, or video input support. M3 catalog estimates use the short-context standard rate rather than calculating tiered or cache pricing.

Security

• No new unsafe code
• No secrets or API keys in diff
• User input validated at boundaries (existing model/driver configuration paths retained)

References: model overview, pricing, Messages compatibility.

[octo-patch]

Thanks for the review. I updated the MiniMax M3 catalog and docs pricing to use the recorded flat input/output rates and removed the unsupported tiered metering path, then pushed a new commit. I ran , , , , and .