The headline: On July 31, two things happened on the same day. OpenAI cut GPT-5.6 Luna by 80% — input from $1 to $0.20 per million tokens, output from $6 to $1.20 — and DeepSeek shipped the official V4-Flash-0731, which reports input at $0.14 and output at $0.28 per million tokens, with a 98% cache-hit discount. By August, analysts were ranking models on “intelligence per dollar” instead of capability. The useful number for engineering teams is neither sticker price nor benchmark score: it is cost per task — and that changes with every workload.
The same-day double punch
The price cut was OpenAI’s first response to sustained pressure from Chinese open-weight models. Luna — the batch-inference model in the GPT-5.6 family — dropped from $1 to $0.20 input and $6 to $1.20 output. The mid-tier Terra dropped 20% ($2.50→$2.00 input, $15→$12 output). Flagship Sol stayed at $5/$30 but gained a “Fast mode” — 2.5× speed at 2× the price, with no intelligence loss. OpenAI also switched the auto-review model inside ChatGPT and Codex CLI from GPT-5.4 to Luna, estimating the cost of that frequent agent workload drops to roughly a tenth. Luna’s input price now matches the previous generation’s nano tier — OpenAI is selling agent-capable models at what used to be small-model prices.
The same day, DeepSeek shipped the official V4-Flash-0731: same architecture and parameter count as the preview, with a pure post-training upgrade focused on agent capabilities — terminal operations, tool use, multi-step execution. Its API pricing: $0.14 input / $0.28 output per million tokens (cache miss), and $0.0028 for cache-hit input — a 98% discount, versus the ~90% most providers offer. DeepSeek reported that on the Artificial Analysis Intelligence Index v4.1 it scores 50, one point behind Luna at 51 — while reporting cost per task about 60% below Luna’s.
The metric changed
The most telling shift is in how the market now evaluates models. Huatai Securities’ August report put it plainly: GPT-5.6’s price cut moves the competition from “capability ranking” to “equivalent intelligence cost.” Their numbers: V4-Flash-0731’s blended price is about $0.06 per million tokens and its average task cost about $0.03 — roughly 65% and 57% below Luna’s. Kimi K3 at 57 represents the capability ceiling among Chinese open-weight models; V4-Flash resets the cost floor in its capability band.
This is exactly the right framing, and it is why token price is the wrong lens. Agent workloads are token-hungry and context-heavy: one coding task can burn hundreds of thousands of tokens across tool calls, retries, and long context. Two models at the same token price can differ by multiples in cost per finished task — and two models at very different token prices can end up nearly equal per task, because one wastes tokens on redundant output and the other does not. DeepSeek reported that post-training cut its output-token consumption by about 12% on the same intelligence-index evaluations. Token efficiency is a cost feature, not a quality footnote.
The dynamic-pricing wrinkle
The price war also made the bill harder to predict. DeepSeek followed its launch with a peak-hour surcharge announcement — roughly doubling prices in peak windows — and the same reporting expects V4-Pro’s API deployment to arrive in early August. Usage-based pricing with peak multipliers means the cost of a task depends on when you run it, not just how many tokens it consumes. For teams, that turns cost modeling from a one-time comparison into an ongoing discipline: measure per-task cost on your own workloads, at your own times, against your own acceptance criteria — which is exactly what a bounded pilot is for.
What the price war doesn’t change
Cheaper tokens do not change the non-price parts of the decision:
- The data path is not priced. Whether your prompts, context, and logs leave your environment is a separate question from what they cost. Teams in regulated contexts still need self-hosted, managed environments regardless of who wins the price war.
- The approval gate is not priced either. The market leader’s Auto Mode removes permission prompts; the cost of re-adding review gates to your workflow is a real number, and it is the same number for every vendor.
- Total cost is workload-shaped. A tool that is cheapest per token may be expensive per task on your specific workload, and the reverse is true too. The only way to know is to run your own tasks end to end and measure.
The take
The July 31 price war is good news for teams in the narrow sense: capable agents are getting cheaper, and the pressure on every vendor’s pricing is structural. The strategic reading is subtler. When token prices converge toward the cost of electricity, the differentiation moves up the stack — to the environment agents run in, the gates around their actions, and the platform that keeps the whole thing auditable. That is where the same checklist applies: prove it on your code, model real per-task cost, know the data path, and keep the review gate in the human workflow. Cheap tokens make bad decisions cheaper to make. They do not make them better.