Frontier model pricing is still falling, and the gap between cheap workhorse tiers and premium tiers is widening — which means the model you shipped with last quarter may no longer be the right economic bet. The blended LLM Price Index hit $4.39 per million tokens on August 1, down from $4.57 in February, according to Model Price Watch. But that average hides a bigger story: some models now cost $0.10 per million input tokens while others cost 200x more. If your product’s unit economics assume last quarter’s pricing, they’re already stale.
Why it matters
Pricing is no longer just a procurement line item — it’s a product decision. Anthropic made Claude Sonnet 5’s introductory pricing ($2 input / $10 output per million tokens) permanent rather than letting it expire, and OpenAI cut GPT-5.6 Sol to $4 input / $20 output per million tokens with a commitment to hold that rate for at least three months. Both moves signal that price competition is sticking, not a temporary promo. For builders, that changes the calculus on whether to lock into one vendor or build routing logic that can shift traffic as prices move.
The numbers that matter right now
- LLM Price Index: $4.39 per million tokens (Aug 1), down 3.9% since February — the market-wide direction is down, but unevenly.
- Gemini 2.5 Flash-Lite: $0.10 input / $0.40 output per million tokens — the current low end for high-volume, low-complexity tasks.
- Claude Sonnet 5: $2 input / $10 output per million tokens, now a permanent rate, not an intro deal.
- GPT-5.6 Sol: cut to $4 input / $20 output per million tokens, guaranteed for at least three months.
- Eval costs: running ScienceAgentBench ranges from $0.19 to $77 per evaluation; TAU-bench Airline runs $0.31 to $180 — meaning benchmarking a new model can cost more than a week of production traffic on a cheap tier.
What this means if you ship product this quarter
None of this is abstract for teams with a live API bill. Here’s the practical read:
- Re-run your cost model before you re-run your roadmap. If you priced a feature around $10-15 per million tokens six months ago, you’re likely overpaying by 3-5x for tasks that don’t need premium reasoning. Route simple classification, extraction, or summarization to sub-$1 tiers like Gemini 2.5 Flash-Lite and reserve premium models for tasks that actually need them.
- Treat “permanent” pricing as a green light to commit, cautiously. Anthropic locking in Sonnet 5’s rate and OpenAI’s three-month floor on GPT-5.6 Sol both reduce near-term repricing risk. That’s a real signal to build cost projections on, but “three months” is not “forever” — bake in a pricing-review checkpoint at the 10-week mark rather than assuming stability indefinitely.
- Build a routing or fallback layer now, not later. New entrants like Qwen3.8-Max, Muse Spark 1.2, and GLM-5.3 are bundling longer context windows at lower prices this month alone. A hardcoded single-model integration means re-engineering every time a cheaper or better option ships. A thin routing layer — even a simple rules-based one keyed on task type and token volume — pays for itself within a quarter for any product doing real volume.
- Budget for evaluation, not just inference. The eval-cost numbers are the sleeper issue here. If comprehensive benchmark runs can cost $77 to $180 per evaluation task, teams that want confidence before switching models need to budget real money and time for that, or they’ll end up “trusting the vibes” of a model swap that breaks something in production. Lightweight, task-specific eval sets — testing on your actual traffic patterns rather than public benchmarks — are usually cheaper and more predictive anyway.
- Reliability still trumps price at the margin. A cheaper model that hallucinates more or times out under load isn’t actually cheaper once you count support tickets and retries. Price should be the second filter, after a task-specific quality bar, not the first.
The honest counter-argument
Chasing every price cut is its own trap. Constant re-benchmarking and re-routing carries engineering overhead, and switching costs — prompt re-tuning, new failure modes, monitoring changes — are real even when token prices drop. For teams without dedicated ML infra staff, sticking with one well-understood model and eating a modest price premium can be the more rational choice than optimizing every dollar of inference spend. The data above says prices are falling; it doesn’t say every team should act on that signal immediately.
Bottom line
The frontier pricing story this quarter isn’t just “AI got cheaper.” It’s that the spread between cheap and premium tiers is now wide enough, and stable enough, that routing decisions are a legitimate product lever — provided you also budget for the evaluation work needed to use that lever safely.