Token prices fall as access tiers rise

What's happening — and why

How it works

Evidence over time

37 supporting · 9 counter — hover or tap a point for detail, click to jump to the row.

Evidence

Trivia

• DeepSeek V2's May 2024 launch at $0.14 per million input tokens reset the entire industry's price-floor expectations for coding-capable models — within 18 months every major frontier vendor had cut per-token prices at least once. No prior software category has experienced a 10-100× cost compression within two years of reaching mainstream adoption, making token deflation structurally unlike any historical software pricing precedent.

• Anthropic's August 2024 prompt-caching launch produced the largest single-day effective price reduction in the corpus (up to 80% off cached input) without any model change — demonstrating that token deflation arrives through infrastructure optimisation (caching, batching) as well as through model-generation cuts, and that the two mechanisms compound. A buyer using both batch (50% off) and caching (up to 80% off cached portions) can reduce effective input cost by 90%+ on workloads with stable system prompts.

• The scissors dynamic — token API prices falling while consumer subscription ceilings rise — is the defining pricing paradox of the 2024-2026 AI market. You.com's 6.7× consumer repricing (from $30 Team to $200 Max) and Character.ai's ad insertion for free users both happened while the underlying token rates kept deflating. The divergence shows that AI products have successfully decoupled human-facing access pricing from the compute costs beneath them.

• Anthropic's June 2026 Fable 5 / Mythos 5 launch is the corpus's clearest proof that token deflation is per-tier, not per-vendor: the company opened a new $10/$50 flagship band — exactly double Opus 4.8's $5/$25 — without cutting or raising any existing tier. Three weeks earlier (2026-05-30) it had rebased Opus itself ~3× downward from the legacy $15/$75. So inside one product line and one month, the established tiers deflated while the 'best-available' ceiling re-inflated 2×, confirming that what falls is the price of a fixed capability level, not the price of the frontier.

• Together AI cut its on-demand H100 rate twice inside a single week ending 2026-06-30 ($4.79 → $3.99/hr), and its $3.09/hr reserved floor now sits BELOW Lambda Labs' $3.99 on-demand rate logged just three weeks earlier — the first time in the corpus that a managed-cluster vendor's reserved price has undercut a peer's on-demand price, evidence that the 'GPU scarcity inflates access' counter-current is breaking down under managed-cluster competition.

• The $3/$15 Claude Sonnet band was the single most durable price constant in the corpus — held flat across four model generations from Claude 3 through Sonnet 4.6 while capability climbed. On 2026-07-06 Sonnet 5 finally bent it, launching at an introductory $2/$10 through Aug 31 before reverting to $3/$15. The anchor didn't break permanently — it dipped — but even a temporary intro cut on the corpus's most stable price is the clearest sign that generational deflation now reaches the tiers vendors most want to keep stable.

• The same 2026-07-06 batch that pushed Sonnet's anchor down also produced the corpus's first frontier-lab token *increase*: Mistral raised Small 4 by 50% on input and 100% on output. Two frontier labs moved their token prices in opposite directions on the same date — the cleanest proof that token deflation is a dominant trend, not a law: a specific under-priced SKU can still be repriced up mid-life.

• DeepInfra broke its own brand on 2026-07-14: the vendor whose reputation in cost-sensitive communities was built on relentless public price CUTS raised dedicated GPU-hour rates 16-32% across the board (H100 +23%, B200 +32%) and lifted its popular DeepSeek-V3.1 tokens from $0.21/$0.79 to $0.25/$0.95 — the corpus's first case of a commodity-inference vendor raising its WHOLE card rather than one under-priced SKU, evidence that even the deflation poster child can be forced up when Blackwell-generation supply and demand tighten.

• xAI priced a new flagship ABOVE its predecessor for the first time on 2026-07-14: grok-4.5 launched at $2/$6 per 1M versus the still-live grok-4.3 at $1.25/$2.50 — reversing xAI's multi-year per-token markdown and keeping the cheaper model live so the premium is reserved for the higher-intelligence tier. Combined with Google's AlphaEvolve 3×-all-in agent surcharge the same day, the frontier ceiling re-inflated via BOTH a higher flagship and an agent multiplier while Together (−15%/−25%) and RunPod kept cutting GPU beneath them.

• The corpus's most under-reported inflation mechanic moves no price at all. Baseten cut its Model APIs card from 11 SKUs to 8 on 2026-07-21, retiring GLM 5.1 ($1.30/$4.30), GLM 5 ($0.95/$3.15), Kimi K2.5 ($0.60/$3.00) and Nemotron 3 Super ($0.30/$0.75) — which lifted the cheapest published cache-input rate on the card from $0.06 to $0.12 without repricing a single surviving model. Groq did the same thing the same day, dropping Llama 4 Scout ($0.11/$0.34) and Qwen3 32B ($0.29/$0.59) with every retained price stable. Delisting the cheap end is a price increase that leaves no diff on any rate.

• Novita delisted 49 of 221 models in eight days — 221 → 196 on 2026-07-21, then 196 → 172 on 2026-07-29 — and in the same window took Flux.1 Kontext Pro from $0.036 to $0.36 per image, a 10x rise that makes it pricier than the ostensibly higher-tier Kontext Max at $0.072. The old rate was confirmed across three independent prior captures, so this is a genuine vendor move on a platform whose entire positioning is "cheapest place to run open-weight models."

• Two vendors ran cuts and raises on the same rate card within eight days. Baseten pruned to 8 SKUs (lifting its floor) on 2026-07-21, then re-expanded to 10 and CUT GLM-5.2 cache input 46% ($0.26 → $0.14) on 2026-07-29. Weights & Biases raised DeepSeek V4-Flash 14x on input and 28x on output ($0.01/$0.01 → $0.14/$0.28) on 2026-07-21, lifting its catalog floor from $0.01 to $0.05 — then cut GLM 5.2 roughly 45% and Kimi K2.7/K2.6 by 25-32% on 2026-07-29. Directional summaries of "the infra layer" no longer survive a week.

• Moonshot ended two years of undercutting in one launch: Kimi K3 arrived on 2026-07-16 at $3.00/$15.00 per 1M — 3.2x the input and 3.75x the output of Kimi K2.6 ($0.95/$4.00) — opening the company's first genuine 5x ladder between its frontier model and its mainstream line. Nine days earlier its most distinctive mechanic was dated for retirement: the moonshot-v1 family, which charged different per-token rates for the SAME weights depending on the context window requested, sunsets 2026-08-31, retiring context length as a billing axis.

• The "everyday" tier was repriced UP by both hyperscale labs in the same week, which no prior generation did. OpenAI's GPT-5.6 Luna launched at $1/$6 per 1M against the outgoing GPT-5.4 mini's $0.75/$4.50, while Sol ($5/$30) and Terra ($2.50/$15) held exactly. Google priced Gemini 3.5 Flash-Lite at $0.30/$2.50 — above Gemini 3.1 Flash-Lite's $0.25/$1.50 — even as Gemini 3.6 Flash cut output 17% ($9.00 → $7.50). Deflation held at the top and the middle; the cheap tier is where both raised.

For buyers

Re-baseline token-cost assumptions at least twice a year — a model you priced six months ago is usually cheaper now, and a newer model may be cheaper and better. Don't extrapolate the deflation to seats: model the workload as raw metered tokens (falling) separately from per-seat access (rising), and pick the cheaper envelope deliberately.

For vendors

To ride deflation you need per-model rate cards that can change without breaking contracts, plus structural discounts (caching, batch) to cut effective price without touching headline rates. To monetise access, a premium seat tier with priority or uncapped usage captures the value that falling tokens leave on the table.

Outlook — what to watch

Expect another leg down: open-weight models (DeepSeek, Llama-class) keep resetting the floor, and caching/batch discounts are spreading. The deflation only stalls if frontier capability stops commoditising — watch whether any vendor can hold a premium token price on a genuinely differentiated model. Access tiers still have room to climb.

FAQ

Related trends

• The $200 prosumer ceiling
• The per-token list price is becoming a base rate times named modifiers
• Credits as the universal currency