In the other direction, the cheap floor went vertical. DeepSeek Pro Latest output $0.396 → $8.00 (+1,920%) — twenty times, in one day. DeepSeek V4.1 Flash input $0.02 → $0.14 (+600%) on the same afternoon its output fell $1.00 → $0.42 (-58.0%). DeepSeek V4 Flash 0731 input +168.0% and output +78.5%; V4 Flash 0423 input -50.7% while its output rose +147.3%. Xiaomi's MiMo-V2.5 output $0.238 → $2.00 (+740%), Kimi Latest output $9.34 → $20.00 (+114%). Seven models were retired outright, six of them Mistral batch lines.

Why the two directions in one day is the whole story. A bill is no longer exposed to a market trend, it is exposed to a line. The Terra half is a launch cut that most invoices will not see for a quarter, because it lands on a model ID the customer never checks; the DeepSeek number is a floor being withdrawn from under everyone who moved onto it in writing. Both happened on 27 October, in the same catalogue, and neither vendor emailed anyone.

The sector's own numbers say the same thing. Anthropic's S-1, filed 26 Oct, pairs $16.3bn of half-year revenue with an $80.7bn loss and a ballooning computing bill (Bloomberg, 26 Oct); the Washington Post's read the same day is that there is not enough money in the world to pay for the AI boom (27 Oct). ExtremeTech, this morning: AI has yet to boost profits for most companies. Apollo's chief economist, 24 Oct: no sign of AI payoff outside the tech sector. The cost is landing on the customer's P&L while the benefit is still a slide.

Concretely, same workload, 200M input and 40M output on Terra: at yesterday's rate card ($2 / $12) that is $880/month; at today's ($1 / $5) it is $400; through the batch door ($0.50 / $2.50) it is $200. Nothing about the work changed — only which line of the catalogue the invoice points at. Re-cost your exact model, get the re-bid email, see my own audit receipts, or book the free clinic.

Today the log went the other way: 21 changes, 12 rises, 8 cuts. The single biggest move is DeepSeek V4 Pro 0423, input $0.2088 → $0.7875 (+277.2%) and output at the same rate. Moonshot's Kimi K3 output $10.80 → $20.00 (+85.2%). Qwen3.6 27B output $2.00 → $3.20 (+60.0%). Against that, GLM Latest output fell 68.1% and Qwen3.5-122B dropped 35% on both fields — so the day is not a wall, it is a squeeze in the mid tier.

Why it matters this week rather than next. Anthropic publicly filed its IPO prospectus today: $16.3bn revenue in the six months to June, up 13.5x, a net loss of $80.7bn (mostly a $72.4bn convertible-note charge), and $430bn locked into cloud compute deals. A company that has to show public investors a margin cannot fund a price war out of a private balance sheet. The 80% cuts were market share bought on someone else's money; the S-1 is the moment that money starts asking for a return. Every earlier brief here found the floor falling. This one finds the mid tier being repriced upward on the same day the receipt for the war got filed.

The buying rule that falls out of it: a bill set on a launch price is a bill on a promotional rate, and a discount adopted to win the market expires when the market is won. Anyone who moved workloads onto the cheap tier in July and never re-read the rate is now on a +277% line. Both directions are live in the same week — which is exactly why the only defensible contract is one with a published-rate clause and a re-bid date.

The whole cheap tier went with them. Of 33 price changes logged today across 470 models — 16 cuts, 12 rises, 5 new — the rises are concentrated exactly where a cost-conscious buyer had just moved. GLM Flash Latest input $0.0113 → $0.06 (+431%), DeepSeek V4 Flash Latest input $0.0044 → $0.018 (+309%) and output +250%, DeepSeek Flash Latest input +180%, DeepSeek V4 Flash 0731 input +267%.

The cuts moved to the batch and committed-volume doors instead. DeepSeek V4.1 Flash (batch) input $0.112 → $0.05 and output $0.336 → $0.15 (-55.4% both sides), DeepSeek Pro Latest input -51.9%, DeepSeek V4 Pro 0423 -48.9% both sides, DeepSeek V3.2 output -47.5%, Kimi Latest output $24 → $14 (-41.7%). Cheap capacity still exists — it is sold to committed volume now, not to the retail tier.

And a new model landed the same morning. DeepSeek V4.1 Flash 1017 entered the log today at $0.60 output; it was the top story on Hacker News this morning (393 points). The cheapest route to the same work changed three times in seventy-two hours.

The industry is now arguing about the same thing. CTech, 17 Oct: "Is the subscription era over? How AI agents are breaking SaaS monetization" — per-seat pricing is being arbitraged by agents. The FT has Anthropic carrying the highest junk credit rating (16 Oct); Reuters has OpenAI's 2026 revenue estimate cut by $20B (17 Oct). When the sellers get repriced, their buyers' contracts are the loosest they will ever be.

Concretely: a team that moved 200M input tokens onto GLM 5.3 Flash on yesterday's numbers was paying $2.26/month this time yesterday and $30.00/month today, same tokens, same work. One day of not looking, and the "saving" is a 13x cost increase. Re-cost your exact model, run your spend, see my own audit receipts, or book the free clinic.

The floor from Monday got taken back. DeepSeek V4.1 Flash input went $0.0238 → $0.15 (+530%) — that is the same model the 13 Oct brief called the cheap floor after a 70.6% cut. DeepSeek Pro Latest output rose +90.9%, GLM 5.3 Flash output +80%. The cuts went to other doors: GLM 5.3 Flash input $0.15 → $0.0113 (-92.5%), DeepSeek Pro Latest output $8.00 → $0.7484 (-90.6%), Meta's Llama 3.3 70B input -83.1%.

The churn is now the headline. The log holds 31 round trips — models that fell more than 20% and climbed straight back — up from 20 yesterday. GLM 5.2 input is the biggest: down 82.5% on 8 Oct, up 933% today. Every round trip, live →

And the money noticed. Today alone: North.cloud raised a $60M Series B to keep runaway cloud and AI costs in check, Ramp closed $1.85B at a $60B valuation and expanded into AI spend management, HubSpot added AI credit caps and pay-as-you-go (15 Oct) and Accenture put a monthly cap on employee Claude usage. The category just got funded — for enterprises. An SMB spending $2k/month still has a spreadsheet, a renewal date, and nobody who will re-cost it against prices that moved 41 times today.

Concretely: 200M input / 40M output on DeepSeek V4.1 Flash cost $4.76/month of input on 15 Oct and $30.00 today — five days after it was called the cheap floor. Re-cost your exact model, run your spend, or book the free clinic.

The cuts did not stop. Moonshot's Kimi Latest input fell 45.9% ($0.536 → $0.29), GLM 5.3 Flash output 44.4%, Kimi K3 input 35.7%, GLM 5.1 27.4% both sides, DeepSeek Flash output 19.2%. Cheap capacity is still there — just not where it was on Monday.

The point is not the direction, it is the churn. The log now holds 20 round trips — a model that fell more than 20% and climbed straight back inside the window. The biggest is Qwen3.8 27B input: down 90% on 11 Oct, up 500% on 15 Oct. A bill costed on a monthly snapshot is now wrong in both directions. See every round trip, live →

Concretely: 200M input / 40M output on gpt-oss-120b cost $14.20/month on 12 Oct, $54.00 on 13 Oct, $14.20 on 14 Oct, and $54.00 again today. Same workload, same week. Re-cost your exact model or run your spend.

The floor still exists, just elsewhere. DeepSeek V4.1 Flash cut input 70.6% ($0.15 → $0.0441) and keeps $0.63 output. GLM 5.3 Flash output fell 25% ($1.20 → $0.90, with its input up 12.5%). Meta also listed Muse Wonder at $20.00 output — a premium tier landing on the same page as four-cent input.

Concretely: a workload of 200M input / 40M output that cost $14.20/month on gpt-oss-120b yesterday costs $54.00 today. The same volume on DeepSeek V4.1 Flash costs $34.02. Nothing about the workload changed. Only the default did.

That is the whole reason a monthly re-cost matters: the models priced for "cheap" are the ones that moved most, and a default set last month is now either bleeding or wrongly routed. Scan your exact model and volume, or run your spend.

That is the whole reason a fixed default is dangerous: two models from the same lab moved in opposite directions on the same night. Whatever your AI bill was costed on is now wrong in one direction or the other — and nobody re-checks it.

The cheap routes still exist: the cheapest input tracked here is now $0.0128/1M tokens (DeepSeek V4 Flash Latest), the cheapest output $0.03 (Mistral Nemo). Run your number, and I'll find which of your lines moved.

The money side stayed blunt the same week: OpenAI's annualised revenue landed roughly $20B below plan (TechCrunch / FT, 8 Oct) and analysts now say AI revenue has to reach $400–600B by 2027 to avoid a spending cliff (Barron's, 9 Oct). Cheaper models, louder ROI questions.

Both halves sit side by side on this page: the invoice is under scrutiny, and the cheaper price already exists. Run your number, and I'll find the cut.

The same week the money side got blunt: analysts now say AI revenue has to reach $400–600B by 2027 to avoid a spending cliff (Barron's, 9 Oct), and the big labs are answering by shipping cheaper tiers (Microsoft's new low-cost decision model landed on Hacker News today, 192pts). Cheaper models, louder ROI questions.

Both halves sit side by side on this page: the invoice is under scrutiny, and the cheaper price already exists. Run your number, and I'll find the cut.

The model market cut again overnight on the same page: DeepSeek V4 Flash output down 97.3%, Z.ai GLM Flash output down 92.6%, Qwen3.5-35B input down 46.7% — logged in the table below.

Both halves of the argument now sit side by side: the spend is under scrutiny, and the cheaper price already exists. Run your number, and I'll find the cut.

Open-weight prices slid again in the same window: DeepSeek V4 Flash input down 66.7%, GLM 5.2 input down 82.5%, DeepSeek Pro input down 62.2% — logged in the table below. If your AI bill was costed on last year's prices, it is now priced against a market that has moved underneath it.

That gap is the whole job: run your number, and I'll find the cut.