The model price list is now the fastest-moving input in an AI budget, and price moves in both directions. A provider can raise list prices mid-quarter — the May 2026 GPT-5.5 price jump was roughly a 2x move on a flagship — and open-weight releases keep grinding the cost-per-task down on the cheap end. Teams that hard-coded the price they remembered found out in their July invoice. This post is the mid-August read.
What moved this cycle
- GPT-5.6 shipped tiered. OpenAI's release lands with separate tiers rather than one flat model price. The practical effect: "the price of GPT-5.6" is no longer a number — it is a table per tier per token profile. Models that used to be quoted by a single in/out pair are now quoted by a matrix.
- DeepSeek V4-P hit GA on August 13. The release notes close out a quarter in which DeepSeek V4 kept taking agentic traffic share on the back of the lowest cost-per-token in its class. If you route any high-volume, function-calling workload, the cheap end of your table moved again.
- Claude moved to usage-based metering this year. Anthropic's list price update landed June 29, and Claude Code's billing mechanics changed on June 15 — the session-relative bill now tracks the model class actually used, which is its own story (see the Claude Code billing post).
- The frontier stopped moving on price. Opus-class and GPT-flagship tokens have been flat in dollar terms since 2023 while everything below them fell. The interesting movement is no longer "which flagship got cheaper" — it is "which tier of which model does my task actually need."
None of these are unique to a provider. The pattern for the quarter is diversification of pricing: tiers, cache differentials, batch rates, context tiers. A price snapshot that only records the headline number is now under-informed by design.
The six columns that make a price readable
The full method — input vs output, cache read vs write, batch discount, context tiers, reasoning-token billing, markup — is written out in the LLM API pricing comparison post. The short version for a monthly tracker is to record, per model:
- $ in and $ out per Mtok, from the post's-published list — not the marketing headline.
- Cache read and cache write multipliers, if the provider supports prompt caching on that path.
- Batch discount and SLA, for models that offer it.
- Context-tier threshold: where the price per token changes above a long-context line.
- Reasoning-token policy: billed or not, visible or opaque.
- Markup/fee: whether the endpoint you route through adds anything on top of the list price. On Mappace's pricing page the answer is zero on token cost, but on your endpoint of record it may not be — it is worth a column.
The monthly diff habit
The tracker pays for itself when the diff happens in front of you, not after the invoice:
- Snapshot the price table on the 1st (the model catalog is the starting point).
- Re-run your top five task pairs (task, typical in/out volume) against the new table.
- Diff against last month's task costs; flag any movement above ~10%.
- Re-check routing rules on every flagged model — the move is either a cut or a re-route, and both are one-line changes.
Twenty minutes per month. The price move that would have been a surprise in August is a line item in a diff you already read. And the single most reported surprise of the past two quarters — a 2x jump on a model you assumed had stable pricing — is exactly the class of event a monthly diff catches before it arrives on an invoice.
The live table, refreshed as providers publish, is in the model catalog. Export the numbers for your top workloads, diff them monthly, and treat a moving price as a routing decision, not a passive cost.