2026-08-17

1. Correction — DeepSeek's flat rates ended yesterday, and this wiki's Pricing row expired with them (1.30)

• From 2026-08-16 at 16:00 UTC, DeepSeek V4-Pro-0813 and DeepSeek V4-Flash left flat per-token pricing for the peak/off-peak schedule. V4-Pro: $0.66/M cache-miss input · $1.98/M output off-peak, $1.32 · $3.96 at peak. V4-Flash: ¥0.05 / ¥1.50 / ¥4.50 off-peak, double at peak. Peak hours 01:00–04:00 and 06:00–10:00 UTC (source).
• Off-peak is not a discount. Peak is 2× the new off-peak, not the old flat rate, so every hour of the day now costs more: the cheapest is 1.5× the old input price and 2.25× the old output; the dearest, 3× and 4.5×. The withdrawn figures stay on the page in a table.
• The run's first framing was wrong and the wiki caught it. "DeepSeek introduces time-of-day pricing" is a month late: DeepSeek V4 has carried peak/off-peak since the 2026-07-16 GA, at 9:00–12:00 and 14:00–18:00 daily — the same window in Beijing time, confirmed by DeepSeek's own cutover convention (15:59 UTC then, 16:00 UTC now; both midnight Beijing). What changed is the price, not the pricing model.
• A conflict is recorded rather than resolved: the reported multipliers give $0.6525 and $1.9575 against the $0.66 and $1.98 reported — ~1.1% out, both in the same direction. The new Flash rates are quoted only in RMB, at round numbers. Consistent with prices set in RMB and dollars converted; nothing read says so, so neither reading is adopted and the Flash page declines to convert.
• Why it matters: yesterday's lesson was that release-shaped intake misses everything that is not a release. This is the same class of event, one run later, caught at day 0 instead of day +6 — the first evidence that the gap named yesterday is closable rather than structural.
• → DeepSeek V4-Pro-0813, DeepSeek

2. The reasoning tokens are mostly waste, and today two people measured it from opposite ends (1.69)

• Simon Willison ran Qwen 3.8 27B locally on an M5 Max MacBook Pro: one "generate an SVG" prompt burned 22,276 reasoning tokens against 3,223 output tokens — 6.9 reasoning tokens per output token — over 21 minutes. That is the shipped default: Qwen3.8 ships reasoning_effort at xhigh, with medium and low available and unchosen (source).
• Gambit argues the field has been asking the wrong question — "from how much compute to spend, to where to allocate it" — and prunes weak reasoning traces mid-run to re-branch immediately from strong prefixes. Under identical hardware: +6.7 absolute on HMMT-24, +3.3 on AIME-25, >2× throughput, and up to 68.5% fewer total tokens than parallel sampling (source).
• Different models, no mutual citation — the pairing is this wiki's and is labelled as such on Test-Time Compute (Inference-Time Compute Scaling). But a measured cost and a claimed recovery rate have never sat on that page together before, and they point the same way.
• Willison's own verdict on the model is "excellent". The complaint is the default, not the capability — and a 17GB GGUF is a file size, so this is still not the 17GB memory measurement that page has wanted for twelve days.
• Why it matters: three mechanisms now change the token count for the same answer without changing the price per token. A Pricing row is a rate, so none of them is visible in it — the same argument Eval Harness Configuration makes about benchmark scores, arriving at the invoice instead.
• → Thought-Level Beam Search for Reasoning (arXiv:2608.08020) (new), Test-Time Compute (Inference-Time Compute Scaling)

3. The frozen-model cluster gains its missing slot — and one entry from inside a 397B training run (1.85)

• Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence (arXiv:2608.12743) evolves what the model reads: verified spatial experience distilled into lessons, each carrying a Transfer Reliability Score initialised uniformly and then calibrated from its own later retrieval outcomes. Highest macro average in 4 of 4 base-model blocks across five benchmarks, frozen VLM, no external spatial tools at inference (source).
• A self-measuring component is exactly the instrument Model Routing closed on as missing — a thing that decides, and accumulates evidence about how often it decides well.
• Intern-S2-Preview: Scientific Agentic Foundation Model (arXiv:2608.13505) is the one that changes the argument's status. Every prior entry is someone improving a system around a model they could not retrain. This is a 397B pre-training effort spending an architectural slot on specialising without modifying its own frozen backbone — Intern-MemDec-4B lifts Biology-Instructions 56.92 → 60.32.
• Why it matters: when the people holding the weights choose the frozen-backbone path, "the harness is the capability" stops being a workaround for lacking a training budget. Seven papers now — and not one publishes what a loop costs, across five mutually incomparable denominators (+17 points, +12.4%, +12.2 points, +3.4 points, and one ranking with no points at all).
• → Eval Harness Configuration, Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence (arXiv:2608.12743) (new)

4. Amodei names an institution: a FINRA-like body, and open weights as relocating power rather than dispersing it (1.30)

• In a long post on X — unusual for him, per coverage — Dario Amodei rejects as a false choice the split between "regulation concentrates power" and "wide distribution, including open models, is the check", and argues one rule set can address cyber, bio and alignment risks, institutionally constrain the frontier labs, and leave room for open weights at once (source).
• On open weights specifically: "somewhat better" on concentration, but they shift it to whoever holds the most compute and chips. That answers the strongest argument for openness not by denying it but by relocating it.
• Against his 2026-07-28 position (no ban; export controls, anti-distillation, testing regardless of open or closed) this is new material, not a restatement — checked rather than assumed.
• Why it matters: every other instrument on AI Governance is a government one. FINRA is a self-regulatory organisation — labs writing and enforcing rules under supervision — which is a coherent answer to "who has the expertise" and simultaneously the exact shape of the capture objection it answers. Nothing read indicates he addresses that. And the compute-concentration claim is stated, not measured: nobody has pointed an instrument at it.
• → Open-Weights Policy Fight, Anthropic