Conversation
…cards DeepSeek-V4.1-Flash shipped 2026-09-10 as `deepseek-flash`. V4 Flash and V4 Flash Vision Exp are retired; their names still work, are served by V4.1 Flash and bill at the Flash price. V4 Pro continues, billing unchanged. GET /models lists exactly deepseek-flash and deepseek-v4-pro. - Default model is deepseek-flash (CLI, jobs, eval runner). Every sub-agent role runs on it, including reviewer, which ran pro. README cites DeepSeek's own changelog for why: on the benchmarks both the 2026-09-10 and 2026-08-13 entries report, Flash leads on Terminal Bench 2.1, NL2Repo, CyberGym and HLE with tools, and trails on HLE without tools. Our own eval has not been re-run on V4.1. - Pricing was still the 2026-08-02 flat card: the 2026-08-16 time-of-use repricing never reached it. The catalog now holds dated cards (flat, V4, V4.1 from 2026-09-10 11:00 UTC, an inferred instant shared with deepseek-cli and dsh-meter) with peak at 2x, weekdays only from 2026-08-22 16:00 UTC. Retired names price as flash. eval/pricing.json carries the same cards, off-peak. - Help and "known:" lists offer only the current names; retired names are still accepted so existing jobs keep validating. - Compaction's summary call now disables thinking. Measured on deepseek-flash: with thinking on, a 512-token cap (the floor on small context budgets) came back 512/512 reasoning and no text, so compaction silently fell back to the emergency summary. With thinking off the same call returned the summary in 165 tokens. The /seek planner keeps thinking and its 8192 cap; one live plan fit. - tests/claims.test.ts pins every published default-model and price claim (README, EVAL.md, eval runner, eval card, roles) to the catalog; each of nine deliberate breaks turned it red. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
DeepSeek-V4.1-Flash shipped 2026-09-10 as
deepseek-flash(changelog 2026-09-10). V4 Flash and V4 Flash Vision Exp are retired; their names are still accepted, served by V4.1 Flash and billed at the Flash price. V4 Pro continues with billing unchanged.What changed
deepseek-flashin the CLI, scheduled jobs and the eval runner. Every built-in sub-agent role runs on it, includingreviewer, which ran pro.eval/pricing.jsoncarries the same cards.deepseek-v4-flashanddeepseek-v4-flash-vision-expare accepted, so existing jobs keep validating, and they price as flash. Help text andknown:lists offer only the current names.deepseek-flash: with thinking on, a 512-token cap (the floor on small context budgets) came back 512/512 reasoning and no text, so compaction fell back to the emergency summary without saying so. With thinking off, the same call returned the summary in 165 tokens. The/seekplanner keeps thinking and its 8192 cap; one live plan fit.tests/claims.test.tspins every published default-model and price claim (README table,--modelline, eval commands in README and EVAL.md, the eval runner's default, eval card, sub-agent roles) to the catalog. Nine deliberate breaks, nine reds.Verification
bun test tests: 150 pass (was 133).dsc -p "Reply with exactly the word: pong" --output-format json->"result":"pong","model":"deepseek-flash", 1083 in / 3 out.No version bump: the V4-Pro GA PR (#1) set the precedent for model-alignment changes.
🤖 Generated with Claude Code