The Codex × DeepSeek model evaluation is recruiting volunteers! Run DeepSeek models through Codex on Distributed Radar to help build the evaluation dataset. Once enough data has been collected, the DeepSeek evaluation results will officially launch on the Codex Radar main site.
Quick links
Some promotional reset cards may expire soon. Open Settings → Usage left → Available resets, then expand Usage limit reset to see the expiry time. Use cards before they expire.
Open Codex Settings → Configuration → Model features → Available reasoning efforts, then select Max. It will appear in model controls for supported models.
What time is safest for distributing reset cards?
Of 31 verifiable resets in Codex Radar history, 19 occurred from 00:00 to 08:59 Beijing time. The site treats 16:00–22:00 as a relatively safer card-distribution window; this is a historical pattern, not a guarantee.
Practical picks derived from the latest distributed test results.
LIVEOnly sustained IQ declines versus the 24/48-hour averages.
LIVELoading update time
Community-tested data · latest successful update shown above. More participants make the data more accurate. Contribute now →
Select one or more configurations to compare their history.
No configurations selected.
This official signal concerns a direct usage reset, not a new banked reset credit.
At Aug 1, 11:32 Beijing time, Tibo confirmed the direct usage reset was complete; no speed window is open.
OriginalFun fact, users use /fast less during the weekend. The weekend is for relaxation, even for the model.
Chinese translation冷知识：用户在周末使用 /fast 的次数更少。周末是用来放松的，模型也不例外。
Phrasesusers use /fast less during the weekend
Model context reading · unofficialThis is a light observation about lower weekend use of /fast. “Even for the model” is personification, not a commitment about limits, resets, or a speed window.
Original@RyanEls4 Codex
Chinese translation@RyanEls4 Codex
Model context reading · unofficialThis is a terse product reference in an app-development conversation, with no reset, limit, or usage commitment.
Original@rezoundous How does it feel
Chinese translation@rezoundous 感觉怎么样？
Model context reading · unofficialThis is an ordinary conversational question. Without earlier explanatory context, it cannot be treated as a reset or quota promise.
Up to 20 runs per desktop view; swipe or use the arrows for older history.
Method: Sol, Terra, and Luna use low reasoning effort with the same count-from-1-to-1024 output task. Standard and Fast each run three independent samples and use the arithmetic mean; order alternates, up to three model pairs run concurrently, and each pair still runs Standard/Fast serially. Every sample uses a fresh app-server and ephemeral thread plus a random nonce. Complete outputs with cached_input_tokens = 0 are preferred. For this run, slow samples and complete cache-mismatched pairs are retained as real same-day experience; incomplete outputs are rejected and rerun. TTFT is delay to the first visible token, TPS is sustained token generation speed, and E2E is total time from submission to the complete result. Headline acceleration values are calculated as mean Standard divided by mean Fast; per-model E2E acceleration uses that model's Standard divided by Fast.