The AI Price War Playbook
You're in. This is the full AI Price War Playbook — bookmark this page.
What just happened
On August 21, OpenAI cut API pricing on GPT-5.6 Sol, its flagship model, from $5/$30 to $4/$20 per million tokens (input/output). The cut is promotional and runs three months, through roughly November 21. OpenAI said the quiet part in its own announcement: the pressure is coming from Anthropic and from Chinese labs.
It's the third pricing move in ten days. Google launched Gemini 3.7 Flash on August 13 at $0.75/$3.75 per million tokens, a 50% introductory discount that expires December 31. And DeepSeek did something nobody else has tried at this scale: since August 16, its V4 models bill by the clock. Peak hours (01:00–04:00 and 06:00–10:00 UTC) cost double the off-peak rate. V4 Pro runs $0.66/$1.98 off-peak and $1.32/$3.96 at peak. That was actually a price increase dressed as a discount, and most people missed it.
Here's why this matters for your bill: the spread between the most and least expensive competent model has never been wider. Claude Fable 5 lists at $10/$50. DeepSeek V4 Flash off-peak is $0.22/$0.66. That is a 75x difference on output tokens. If you're running everything through one frontier model out of habit, you are donating money to a lab that doesn't need it.
The August 2026 price map
| Model | Input /1M | Output /1M | Watch out for |
|---|---|---|---|
| Claude Fable 5 (Anthropic) | $10.00 | $50.00 | Frontier quality, frontier price |
| Claude Opus 5 (Anthropic) | $5.00 | $25.00 | — |
| GPT-5.6 Sol (OpenAI) | $4.00 | $20.00 | Promo ends ~Nov 21; list is $5/$30 |
| Qwen3.8-Max hosted (Alibaba) | $2.00 | $6.00 | Open weights (text only) if you self-host |
| DeepSeek V4 Pro (peak) | $1.32 | $3.96 | Peak = 01:00–04:00, 06:00–10:00 UTC |
| Gemini 3.7 Flash (Google) | $0.75 | $3.75 | Promo ends Dec 31 |
| DeepSeek V4 Pro (off-peak) | $0.66 | $1.98 | 17h/day window |
| DeepSeek V4 Flash (off-peak) | $0.22 | $0.66 | Cached input drops to $0.007 |
Prices per million tokens, standard context, as reported August 21–24, 2026. Verify against the provider's live rate card before you commit budget; promos move.
The framework: route by cost per task, not cost per token
Per-token prices are how labs advertise. Cost per task is what you actually pay. The math is one line:
cost per task = (input tokens × input rate + output tokens × output rate) ÷ 1,000,000
Take a real agent run: 60,000 input tokens (context, files, tool results) and 8,000 output tokens. Same task, five models:
| Model | Cost per run | 1,000 runs/month |
|---|---|---|
| Claude Fable 5 | $1.00 | $1,000 |
| Claude Opus 5 | $0.50 | $500 |
| GPT-5.6 Sol (promo) | $0.40 | $400 |
| Gemini 3.7 Flash | $0.075 | $75 |
| DeepSeek V4 Pro (off-peak) | $0.055 | $55 |
Now the routing rule. Sort every recurring workload into one of four tiers:
Tier 1 — frontier required. Multi-step agentic work on unfamiliar code, high-stakes drafting where a wrong answer costs real money, hard reasoning chains. Pay for Fable 5, Opus 5, or Sol here and don't feel bad about it. A $1 run that saves an hour is the cheapest labor you will ever buy.
Tier 2 — strong but standard. Long-document summarization, code generation with tests that catch mistakes, structured extraction from messy input. Gemini 3.7 Flash and DeepSeek V4 Pro were built for exactly this tier, and the benchmark gap to frontier models on these tasks is small.
Tier 3 — bulk and batch. Classification, tagging, eval runs, dataset cleanup, anything you run 10,000 times. DeepSeek V4 Flash off-peak, or self-hosted Qwen3.8-Max if you have the GPUs. At $0.22/$0.66 the constraint stops being money and becomes rate limits.
Tier 4 — latency critical. User-facing autocomplete and chat. Price matters less than time to first token; test the fast variants (OpenAI is previewing a Sol Ultrafast on Cerebras hardware that it claims runs up to 14x faster).
The gate between tiers is a quality floor, not a feeling. Build a golden set of 30 real tasks from your own logs. Run them through the cheaper model. Grade blind against the frontier output. If the cheap model matches or beats frontier on 95% of your golden set for that workload, route it down. If it doesn't, you just bought certainty for the price of an experiment. This router outlives this week's prices: rates change, the method doesn't.
The off-peak shift
DeepSeek's peak window is 7 hours a day: 01:00–04:00 and 06:00–10:00 UTC. Everything else is half price. Batch work has no business running at peak.
The recipe: queue non-interactive jobs (evals, backfills, classification sweeps, report generation) and release the queue inside the 10:00–01:00 UTC window. A cron line like 0 11 * * * (11:00 UTC daily) clears the morning peak with an hour of margin. If your stack is on the US East coast, note that 9pm–11pm ET falls inside the 01:00–04:00 UTC peak block — the evening batch run that feels off-peak to you is billed at double.
One more lever most teams ignore: cached input. DeepSeek bills cache hits at $0.007–$0.044 per million depending on model and time. If your pipeline re-sends the same system prompt and context on every call, restructure so the static prefix stays identical byte for byte. Cache pricing only helps when the prefix matches exactly.
The promo cliff calendar
Two dates belong in your budget file right now: November 21, 2026 (GPT-5.6 Sol reverts to $5/$30) and December 31, 2026 (Gemini 3.7 Flash promo ends). If you sign a client SOW, set a retainer, or forecast a quarter using promo rates, you have built a 25–50% cost jump into month four and called it a plan. Budget at list price, treat the promo delta as margin, and set a calendar reminder two weeks before each cliff to re-run your router math. Labs are counting on you to anchor on the promo number and stay put when it expires. Don't.
Copy-paste: the LLM cost audit prompt
Paste this into any capable model along with a usage export (CSV from your provider dashboard, or even a rough list of your workloads with volumes). It is written to be used exactly as is.
You are an LLM cost auditor. Your job is to find the cheapest model mix that meets my quality bar, using only the data I give you. Do not invent usage numbers, do not assume workloads I have not listed, and flag any place where my data is too thin to support a recommendation. Input: I will provide (1) a list or export of my LLM workloads with, where known, monthly call volume, average input tokens, average output tokens, current model, and whether the task is interactive or batch; (2) the price table I am working from. Method: For each workload, compute current monthly cost as (input tokens x input rate + output tokens x output rate) x volume / 1,000,000. Then classify it into one of four tiers: frontier-required (multi-step reasoning where errors are expensive), strong-but-standard (summarization, extraction, routine codegen), bulk-batch (classification, tagging, evals at volume), or latency-critical (user-facing, speed first). Propose the cheapest model per tier from my price table, compute the new monthly cost, and show the delta. For batch workloads on time-of-day billing, assume off-peak rates and say so. Boundaries: Never recommend switching a frontier-required workload down on cost alone; instead specify the eval I should run first (golden set size, pass threshold). If two models are within 15% on cost, prefer the one I already use, because migration has its own cost. Definition of done: A table with one row per workload showing current model, current monthly cost, proposed model, proposed monthly cost, savings, and a risk note. Below the table, the three largest savings opportunities ranked, each with the single next action required to capture it. If total projected savings are under 10%, say plainly that switching is not worth the effort this quarter. Escalation: If my usage data lacks token counts, stop and give me the exact export or logging step I need to get them before you estimate anything.
The model-switch checklist
Before any workload moves to a cheaper model, run this. A migration is done when every box is checked, not when the first output looks fine.
1. Golden set built: 30+ real examples pulled from production logs, not toy prompts.
2. Blind grade complete: cheaper model hits ≥95% of incumbent quality on the golden set, graded without knowing which output came from which model.
3. Failure cases read: you looked at every miss individually and none of them is catastrophic (silent data corruption, confident wrong numbers, policy violations).
4. Prompt re-tuned: the new model got its own prompt pass; prompts do not transfer cleanly between labs.
5. Rollback trigger written: a specific metric and threshold that reverts the route automatically (e.g., task success below 92% over 200 runs).
6. Promo cliff noted: if the new price is promotional, the reversion date is in the budget and the re-evaluation is on the calendar.
Common failure modes and the exact fix
Anchoring on promo pricing. You model Q4 on Sol at $4/$20; on November 22 you're at $5/$30. Fix: budget at list, book promo savings as variance, re-run the router two weeks before each cliff.
Comparing per-token instead of per-task. A model with cheap tokens that writes verbose output can cost more per job. Fix: always multiply by your real token counts; a chatty model's output tokens are where bills explode.
Switching without evals. The first week looks fine, then edge cases pile up in support tickets. Fix: the golden set gate above; 30 examples takes an afternoon and prevents a quarter of quiet damage.
Running batch at peak. Your 9pm ET cron job lands in DeepSeek's 01:00 UTC peak block at 2x. Fix: shift batch to the 10:00–01:00 UTC window; one cron edit is a permanent 50% cut on that workload.
Ignoring cached input rates. Rebuilt prompts on every call mean you pay full input price for the same instructions thousands of times. Fix: freeze the static prefix byte for byte and put variable content at the end.
Self-hosting for pride. Qwen3.8-Max weights are free; the 2.4T-parameter cluster to serve them is not. Fix: price the GPUs, ops time, and utilization honestly against $2/$6 hosted before you build.
This playbook is free, like every daily drop. If it saves you real money, you can fuel the daily drops here.