feat: deepseek-flash everywhere, and web_search is gone

[lroolle]

Follows #26. That PR moved the CLI's default to deepseek-flash. This one moves everything else: the docs, the site, and the free tier. It also acts on the two things #26 found and left open.

Web search: removed upstream, so we refuse it

DeepSeek removed web_search from the Responses API with V4.1 Flash on 2026-09-10. Its Responses API guide now lists web_search among the built-in tools that are ignored, and the response.web_search_call.* events are gone. A probe on 2026-09-18 matched: the tool was accepted, but no search ran, and the request billed 21 input tokens where a search used to bill about 40K.

• CLI: respond --web-search is hidden and exits 1 with that reason. It sends nothing.
• Gateway: refuses the tool with a 400 that gives the reason. DailySearches, searchInputAllowance and the search admission path are deleted. They priced searches that could no longer happen.
• Copy: every page that offered search now says it was removed, with the date. TASTE.md records why the 2026-08-07 decision expired.

Models

• Everywhere: deepseek-flash is the default and the recommended pick. Pro stays callable and correctly priced. Nothing ranks it above flash any more. The claim that flash is ahead cites DeepSeek's changelog and release note. The bench page compares only benchmarks that carry the same name on both DeepSeek charts, and it shows HLE-no-tools, where flash is behind.
• Gateway:
  • DSGATE_MODEL defaults to deepseek-flash.
  • resolve() follows the Anthropic API guide exactly: claude-opus* goes to pro, and every other name goes to flash.
  • The retired flash names are the same model on both sides of the check.
  • A DSGATE_MODEL from before the rename is served under the canonical name.
  • This also fixes a live bug: freeseek's /models returns "data": [] today. It is configured as deepseek-v4-flash, and upstream lists only deepseek-flash.
• History stays history: the August flat card, the V4 cost-per-task figures on /economics and the V4 launch table are kept and labelled. They are not relabelled as V4.1.

Pricing instant

V41At moves from the inferred 11:00 UTC to 04:00 UTC on 2026-09-10, the instant DeepSeek's release note gives. The change is in the CLI and gateway gates, the tests (04:00 is also the end of a peak window) and the copy.

Tripwires

Each tripwire below was broken on purpose, went red, and was restored to green.

• internal/cli/claims_test.go: every "(default)" model in README, AGENTS, SKILL and llms.txt must equal deepseek.ModelFlash, and no doc may offer respond --web-search.
• gateway/internal/server/claims_test.go: the DSGATE_MODEL default must be a current name, the index, terms and economics pages must name it, and no page may promise web search.
• site/build.py check_model_claims: a section fails if it names a retired flash name without deepseek-flash, or has live web-search copy, unless it is marked <!-- claim:historical -->.

Verified

• make check is green: it builds, runs the tests, price-check, the gateway with -race, and site-check.
• go test -race ./... passes.
• make cover-gate reports 72.9%.
• Headless Chromium at 390px and 1280px: no page scrolls sideways.

Found, not done

OpenCode Zen still lists deepseek-v4-flash-free, but answers every request with Model is unavailable. Every free-tier chat therefore pays one refused round trip, and its prompt still reaches Zen. The gateway README and DESIGN now say to unset OPENCODE_API_KEY until Zen serves a free V4.1 model. Changing that on the box is an operator decision.

🤖 Generated with Claude Code