The August 2026 AI API Price War: OpenAI Cuts, DeepSeek V4-0731, and the New Cost Floor
OpenAI cut GPT-5.6 Terra and Luna, DeepSeek upgraded V4 Flash, Google launched Gemini 3.6, and Claude Sonnet 5 entered with an introductory price. Here is what the August 2026 API price war means for developers.
The August 2026 AI API price war is real, but it is not a uniform race to zero. OpenAI made the most aggressive permanent move by cutting GPT-5.6 Luna from $1/$6 to $0.20/$1.20 per million input/output tokens. DeepSeek shipped a stronger V4 Flash checkpoint at the same base price. Google lowered premium Flash output cost with Gemini 3.6, while Anthropic is using a temporary Sonnet 5 introduction price rather than a permanent cut.
The result is a market with three different pricing strategies: permanent price compression, time-limited acquisition pricing, and performance upgrades at an unchanged token rate.
What Actually Changed
| Provider and model | Previous reference | Current price | Change |
|---|---|---|---|
| OpenAI GPT-5.6 Terra | $2.50 input / $15 output | $2.00 / $12.00 | 20% lower |
| OpenAI GPT-5.6 Luna | $1.00 input / $6 output | $0.20 / $1.20 | 80% lower |
| Google Gemini 3.5 -> 3.6 Flash | $1.50 input / $9 output | $1.50 / $7.50 | Output 16.7% lower |
| Anthropic Sonnet 4.6 -> Sonnet 5 | $3 input / $15 output | $2 / $10 through Aug 31 | 33.3% lower, temporary |
| xAI Grok 4.5 cached input | $0.50 cached input | $0.30 cached input | 40% lower |
| Mistral Small 4 | $0.10 input / $0.30 output | $0.15 / $0.60 | Input +50%, output +100% |
OpenAI’s current API pricing confirms the new Terra and Luna rates. Sol remains at $5/$30, so the family is no longer a neat 5:2.5:1 price ladder. Luna now costs only 4% of Sol on both input and output while retaining the same published 1.05M context and 128K maximum output.
Google’s Gemini pricing lists Gemini 3.6 Flash at $1.50/M input, $0.15/M cached input, and $7.50/M output. Anthropic’s Claude pricing explicitly schedules Sonnet 5 to move from $2/$10 to $3/$15 on September 1, so August forecasts should not be copied into Q4 budgets.
Did DeepSeek Release a New Model?
DeepSeek released a new checkpoint, but did not create a new public API model ID. On July 31, deepseek-v4-flash began serving DeepSeek-V4-Flash-0731 in public beta. DeepSeek says the update keeps the same architecture and model size while improving agent behavior, Responses API support, and Codex-style workflows. The current token price remains $0.0028/M cached input, $0.14/M cache-miss input, and $0.28/M output. See the DeepSeek changelog and official price table.
There is an important future cost change. DeepSeek has announced 2x billing during Beijing-time peak windows of 09:00-12:00 and 14:00-18:00, but the pricing page still says the effective date is to be determined. It would be wrong to claim the multiplier is already active, and equally risky to omit it from capacity planning.
If that policy starts unchanged, a V4 Flash workload costing $0.42 at base rates would cost $0.84 when fully executed in the peak window. That is still low, but it changes the value of batch scheduling and multi-provider failover.
The New Cost Floor
For a normalized workload of 2M cache-miss input tokens plus 500K output tokens, standard short-context prices produce this bill:
| Model | Cost | Relative to GPT-5.6 Sol |
|---|---|---|
| DeepSeek V4 Flash | $0.42 | 1.7% |
| Xiaomi MiMo-V2.5 | $0.42 | 1.7% |
| Mistral Small 4 | $0.60 | 2.4% |
| GPT-5.6 Luna | $1.00 | 4.0% |
| MiniMax M3 | $1.20 | 4.8% |
| Gemini 3.5 Flash-Lite | $1.85 | 7.4% |
| GLM-5.2 | $5.00 | 20.0% |
| Grok 4.5 | $7.00 | 28.0% |
| Claude Sonnet 5 | $9.00 | 36.0% |
| GPT-5.6 Terra | $10.00 | 40.0% |
| Kimi K3 | $13.50 | 54.0% |
| GPT-5.6 Sol | $25.00 | 100% |
This is not a benchmark ranking. A cheaper model can cost more per successful task if it retries, produces longer reasoning traces, or escalates frequently. A premium model can be economical when one correct call replaces several failed low-cost calls.
Why OpenAI’s Luna Cut Matters
DeepSeek and MiMo still set the lowest uncached text-token floor, especially with cache hits. Luna’s significance is different: it brings a current OpenAI model with a 1.05M context window into the same budget band as MiniMax and the newest Flash-Lite models.
That makes an OpenAI-only router more practical:
- Send extraction, classification, search triage, and routine changes to Luna.
- Escalate normal production coding and document work to Terra.
- Reserve Sol for high-impact or failed requests.
- Use Batch or Flex for asynchronous traffic and Fast mode only when latency justifies 2x token rates.
OpenAI renamed Priority processing to Fast mode on July 30. The service-tier name changed; the pricing distinction remains.
The Price War Is Also About Caching
Headline input/output rates hide the strongest discounts:
- DeepSeek V4 Flash cached input is $0.0028/M, 2% of its cache-miss rate.
- GPT-5.6 models charge 10% of input for cache reads, but explicit cache writes cost 1.25x input.
- Mistral lists a 90% discount for cached input.
- Anthropic cache reads cost 10% of base input, while 5-minute and 1-hour writes cost 1.25x and 2x.
- Kimi K3 uses automatic context caching at $0.30/M cached input versus $3/M cache-miss input.
For long-running agents, cache-write policy and prefix stability can matter more than a 10% headline price difference.
What Developers Should Do Now
Do not migrate on price alone. Run a fixed evaluation set and record cost per successful task, p95 latency, retry rate, output-token use, cache-hit ratio, and tool-call fees. Then use at least two routes: a low-cost default and a quality escalation path.
Also treat promotional prices as dated configuration. Sonnet 5 changes on September 1, Qwen3.7 alias discounts have no published end date, and DeepSeek’s peak multiplier has no published start date. Each needs an operational alert rather than a static spreadsheet assumption.
Use the AI Model Pricing Calculator to model your own traffic, and keep the August API pricing comparison open for the current provider table.
Bottom Line
OpenAI did cut prices, and Luna is the biggest move: an 80% reduction that turns it into a credible high-volume route. DeepSeek’s new V4-Flash-0731 checkpoint holds the lowest base price without introducing a new API ID. Google and Anthropic are applying narrower pressure through a lower-output-price Flash model and a temporary Sonnet launch rate.
The market is not converging on one winner. It is converging on multi-model routing, where price, cache behavior, context thresholds, temporary promotions, and success rate are evaluated together.