Three major AI API pricing moves hit in 30 days. DeepSeek quadrupled output costs overnight. Anthropic froze Claude Sonnet 5 at its intro price permanently. OpenAI slashed Luna by 80%. If you have not rechecked your model routing this month, you are either overpaying or one billing cycle away from sticker shock.
DeepSeek V4 Just Proved Why Single-Vendor Dependency Is Dangerous #
On August 16, DeepSeek switched V4-Pro and V4-Flash from flat pricing to a peak/off-peak structure — and the numbers are ugly. V4-Pro output went from $0.87 per million tokens to $3.96 at peak (US and EU business hours) and $1.98 off-peak. That is a 355% spike at peak, more than double even in the cheap window. V4-Flash output moved from $0.28 to $1.32 at peak.
Teams burning 10 million output tokens per day are looking at bills that went from roughly $8,700 to $39,600 per day at peak. DeepSeek gave a week’s notice on August 6 and cited demand overwhelming capacity. Reasonable — but developers who built cost-critical workloads around DeepSeek’s rock-bottom rates discovered what happens when a pricing moat turns out to be a temporary growth strategy, not a sustainable business model.
The lesson is not to stop using DeepSeek. Off-peak V4-Flash remains cheap by frontier standards. The lesson is: never build budget forecasts assuming a single provider’s price floor will hold. Add a routing layer. This week’s spike is the argument you needed to finally do it.
The Good News: Claude Sonnet 5 Just Got Permanently Affordable #
Anthropic launched Claude Sonnet 5 in June at $2 per million input tokens and $10 per million output — explicitly framed as introductory pricing that would increase to $3/$15 on September 1. On August 10, Anthropic reversed course: the $2/$10 rate is now permanent. The planned 50% hike is cancelled.
For developers, this is the clearest win of the month. Claude Sonnet 5 is a genuinely capable model — strong coding, 200K context, solid agentic workflows. At $2/$10 with no expiry, it undercuts GPT-5.6 Terra ($2/$12) on output while holding comparable quality for most tasks. If you are still on Sonnet 4.6 at $3/$15, switch now. The cost argument is settled.
OpenAI’s Luna Is the New Budget Workhorse #
On July 30, OpenAI permanently cut GPT-5.6 Luna from $1/$6 to $0.20/$1.20 per million tokens — 80% off. Terra dropped 20% to $2/$12. Sol held at $5/$30 on OpenAI direct, though it is available at roughly $2.50/$15 via OpenRouter’s exclusive discount for approximately one month.
Luna is worth building into your toolkit for tasks that do not need heavy reasoning: classification, structured extraction, summarization, high-volume QA pipelines. At $0.20 input and $1.20 output, it matches DeepSeek off-peak prices with better latency predictability. Route your lighter workloads here. Stop paying Sol rates for tasks Luna handles fine.
August 2026 AI API Pricing Snapshot #
| Model | Input ($/M) | Output ($/M) | Stability |
|---|---|---|---|
| GPT-5.6 Sol (OpenRouter) | $2.50 | $15.00 | ~1 month promo | | GPT-5.6 Sol (OpenAI direct) | $5.00 | $30.00 | Stable | | GPT-5.6 Terra | $2.00 | $12.00 | Permanent cut | | GPT-5.6 Luna | $0.20 | $1.20 | Permanent cut | | Claude Opus 5 | $5.00 | $25.00 | Stable | | Claude Sonnet 5 | $2.00 | $10.00 | Permanently locked | | Gemini 3.7 Flash | $0.75 | $3.75 | Doubles Jan 1, 2027 | | DeepSeek V4-Pro | $0.44 | $1.98–$3.96 | Peak/off-peak live | | DeepSeek V4-Flash | $0.14 | $0.66–$1.32 | Peak/off-peak live |
The Routing Strategy That Actually Makes Sense Now #
Stop routing everything to one model. The August 2026 pricing landscape makes tiered routing the practical default:
Complex reasoning and agents: Claude Sonnet 5 at $2/$10 (permanent) or Sol via OpenRouter while the discount holdsHigh-volume, simpler tasks: GPT-5.6 Luna at $0.20/$1.20 or Gemini 3.7 Flash at $0.75/$3.75DeepSeek: Off-peak batch jobs only — keep it out of production traffic during US/EU business hoursThe Gemini clock:$0.75 input doubles on January 1, 2027 — account for that in Q4 planning
Tools like OpenRouter and LiteLLM make multi-provider routing a configuration problem, not a code problem. If you are paying $5/$30 for Sol on OpenAI direct and routing a mix of task difficulties there, you are leaving money on the table. Segment by complexity. The infrastructure to do it already exists.
The Only Safe Assumption Is That Prices Will Change #
AI token prices are not stable. Gemini Flash doubles in January. DeepSeek just proved that introductory pricing is not a permanent competitive advantage. Even the Anthropic Sonnet 5 permanency announcement came 21 days before the scheduled hike. The pricing landscape you are working with today will look different in 90 days.
Build applications that are model-agnostic at the API boundary. The providers earning loyalty right now are those with predictable pricing signals — and Anthropic locking Sonnet 5 is the clearest developer-friendly move of the month. Check your Anthropic pricing docs and swap Sonnet 4.6 workloads before you forget about it. Review your DeepSeek usage and decide if peak-hour exposure is acceptable.
The AI API price war is deflationary in the long run — but only if you are positioned to take advantage of it. Diversify providers, automate routing, and audit your bills every 30 days. That is the strategy now.