Price Index Delta: Seven Rows Move Again, This Time With No Pattern

The seventh DeAI Price Index observation is dated 2026-11-02 UTC, one week after issue six. Comparing OpenRouter's public catalog against the 2026-10-26 snapshot, seven of seventeen tracked rows moved — and this time they moved in opposite directions with no shared pattern, the direct opposite of issue six's synchronized reversion. The headline: deepseek/deepseek-v4-flash-0731 collapsed to $0.119/$0.238 per million tokens (input -73%, output -82%), erasing last week's reversion to $0.44/$1.32 after exactly one observation. The catalog itself shrank to 452 models from 467.

Key facts

• DeepSeek V4 Flash 0731: $0.44/$1.32 → $0.119/$0.238 (input -73%, output -82%) — one week after reverting to $0.44/$1.32, the row is back near the lows of 2026-10-05 ($0.0152/$1.28) and 2026-10-12 ($0.016/$0.64). Four observations, four different rows; no shape has ever held two Mondays. Observed on the OpenRouter models API.
• GLM-5.3 Flash: $0.15/$0.50 → $0.0285/$0.0923 (input -81%, output -82%) — a fourth distinct row in four observations: $0.09/$0.30 (09-21) → $0.15/$0.50 (10-05) → $0.0232/$1.20 (10-12) → $0.15/$0.50 (10-26) → now $0.0285/$0.0923, the cheapest this row has ever been observed.
• DeepSeek V4.1 Flash: $0.02/$1.00 → $0.045/$0.50 (input +125%, output -50%) — last week's watch-item shape lasted exactly one observation. The input leg is back above $0.02 but still below the $0.30 official peak rate (re-verified 2026-10-05); the row has now failed the two-observation test it needed.
• DeepSeek V4 Pro: $0.95526/$1.91052 → $1.60/$3.20 (both +67%) — the issue-5/issue-6 anomaly resolved last week, then jumped 67% past its old level. The 2:1 output/input ratio is the one stable feature of this row.
• GLM-5.3 (full-size): output +157% to $3.60 (input +14% to $0.08) — after issue six's -71%, the output leg has now swung from $4.80 → $1.40 → $3.60 in three observations.
• Kimi K3: $0.152/$20.00 → $0.192/$15.00 (input +26%, output -25%) — the 131× spread narrowed to about 78×. Qwen3.6-35B output slipped 5% to $0.90. Ten rows held flat, including both closed anchors (claude-sonnet-4.6 $3.00/$15.00, gpt-5.5 $5.00/$30.00). Hermes 4 70B is absent a seventh consecutive observation.

What happened

Issue six's diff had a single story line: seven movers, all reverting. This week's diff is the counter-case — the same count of movers, and not one of them matches another's direction. deepseek-v4-flash-0731 and glm-5.3-flash collapsed roughly 80% while deepseek-v4-pro rose 67% and glm-5.3 output rose 157%. If last week demonstrated that this catalog can revert in unison, this week demonstrates that the following week it can scatter.

The two-observation discipline did real work this cycle. deepseek-v4.1-flash's $0.02/$1.00 shape — flagged last week as "re-check next Monday before treating any of it as a cut" — failed the test: the row moved again to $0.045/$0.50, a shape that matches no published DeepSeek card either. Conversely, the ten flat rows now include four that have not moved in at least three observations (deepseek-v4-flash 0423 at $0.091/$0.182 through two, llama-4-maverick at $0.1875/$0.6525, gemini-2.5-flash at $0.30/$2.50, kimi-k2.5 at $0.50/$2.50, qwen3-coder at $0.30/$1.00) — the basket's slowly stabilizing core. The catalog also contracted by 15 models week-over-week (467 → 452), the first shrink the index has recorded; a full de-listing diff is a follow-up exercise beyond this story's basket.

Why it matters

• Last week's "reversion cluster" is not the new normal — volatility is. The strongest single conclusion from seven observations is that this catalog's volatile rows produce a different story every Monday, and any one Monday's pattern (synchronized or scattered) is a bad predictor of the next. The index's two-observation rule caught another false signal this week; a builder who re-priced deepseek-v4-flash-0731 workloads off last Monday's $0.44 row paid 3.7× more per input token than this Monday's row, and one who re-priced off this Monday's $0.119 row against last week's table would have the wrong number entirely.
• The cheapest observed rows keep getting cheaper — with no floor in sight. GLM-5.3 Flash at $0.0285/$0.0923 and deepseek-v4-flash-0731 at $0.119/$0.238 are the lowest rows either model has carried in this index. Neither holds a first-party explanation, and both sit below DeepSeek's own official peak card for the deepseek row — the catalog remains unverified for tracker purposes until a first-party page confirms.
• The deepseek-v4.1-flash saga enters its third week with zero confirmed shapes. $0.30/$1.20 for six Mondays → $0.02/$1.00 → $0.045/$0.50. The family is clearly being restructured around the new 1017 checkpoint (which the index will basket-track once it settles), but after three distinct shapes in three observations, no single number in this family can yet be called "the price."

Background

The DeAI Price Index is a weekly, dated, append-only observation of list pricing — it records what a catalog listed on a date, and never rewrites prior issues. Issue one established the basket and methodology: OpenRouter's GET /api/v1/models prompt/completion fields converted to dollars per million tokens, batch rows excluded, plus first-party anchors. Issue six resolved a three-issue artifact cluster into synchronized reversions; this issue scatters them. Across seven observations the pattern that has held is structural: catalog rows for the same model move by tens of percent in single weeks with no first-party announcement, which is why the tracker grades aggregator pricing unverified regardless of how many Mondays it holds.

A process note, per the standing gap report: the automated DIPT pricing fetcher again produced no machine-readable OpenRouter candidates in this week's sweep (data/tracker/resolved_2026-11-02.json carries two third-party d_pricing claims — a Chutes subscription-terms change and a Venice model-card listing — neither with a machine-readable pricing source configured), so the snapshot was fetched directly for this job using the documented Issue-1 methodology — the sixth week running without a wired OpenRouter adapter. The raw pull is dated and archived (data/tracker/raw_openrouter_models_2026-11-02.json, 452 models fetched live 2026-11-02 ~12:10 EET; basket extraction in pricing_fetch_2026-11-02.json); nothing here derives from provider marketing claims.

What's next

The next observation is scheduled for Monday, 2026-11-09. Watch items: whether the deepseek-v4-flash-0731 and glm-5.3-flash lows hold a second Monday (their first real two-observation test), whether deepseek-v4-pro's 67% jump sticks or reverts, whether glm-5.3's output leg keeps oscillating around the $1.40–$4.80 band, whether kimi-k3's spread keeps narrowing, whether the 15-model catalog contraction is pruning or churn, and whether Hermes 4 70B returns to the catalog. The machine-readable snapshot updates at /data/prices.json alongside this story, with the 2026-10-26 snapshot preserved in the file's archive chain.