AI Token Price War: DeepSeek V4.1 Flash vs GPT-6.1 Sol

Two cheap models, one message

In roughly three weeks, two of the most closely watched AI labs shipped models built on the same pitch: near-frontier coding and agent performance for a fraction of the usual per-token cost. DeepSeek released V4.1 Flash on September 10, 2026 1. OpenAI followed with GPT-6.1 Sol, dated September 29 5 and unveiled at DevDay 2026 a week after GPT-6 Sol 4. Third-party hosting followed soon after. A Luminal-served build of DeepSeek-V4.1-Flash appeared on LLM Gateway on October 4 5.

For developers who run coding agents and long multi-step workflows, the timing matters. The cost of frontier intelligence is no longer just a line item. It is the main thing the labs are competing on.

DeepSeek's offer: open weights and very cheap cached input

V4.1 Flash is described as a 552B-parameter open-weight model. Only a small slice of those parameters, reportedly around 8B to 16B, is active at a time 1. It ships with a one-million-token context window and native vision 2. Users report speeds above 150 tokens per second and performance comparable to top models on coding and agent tasks 1.

The pricing is where it stands out. Uncached input costs $0.15 per million tokens off-peak and $0.30 at peak. Output runs $0.60 off-peak and $1.20 at peak 23. Cached input is far cheaper: $0.003 off-peak and $0.006 at peak per million tokens 23. The larger V4 Pro sits at $0.66 for uncached input and $1.98 for output off-peak, also doubling at peak 2.

Sources differ slightly on the numbers. Social posts describe Flash's API as "starting at $0.30" per million input tokens 1, which is actually the peak cache-miss rate. One pricing tracker's competitor-comparison table lists Flash at $0.14 input and $0.28 output 3. That same site's price history shows those figures were only in effect until September 9 3. So the comparison table appears to be stale. Output has roughly doubled since then, while input moved up only slightly.

The gap between cache hits and misses is the detail that matters most in practice. As one analysis notes, stable prompt prefixes, reused repository context and repeated agent instructions become economically important at a 50x price difference 2. One estimate puts a 100-million-token-per-day workload at about $610 a month off-peak, but that assumes an 80% cache-hit ratio 3.

This also puts the viral claims in context. One developer said they cancelled Claude Code and Codex subscriptions and now spend $10 a month, with $10 buying "~2B tokens" 1. That works out to about half a cent per million tokens. It is only plausible if nearly all of the traffic is cached input. Uncached usage at those volumes would cost many times more. The enthusiasm is real, and the low cost is real too. It depends heavily on how a workload is structured.

OpenAI's answer: the same price cut, without open weights

OpenAI's response does not try to match DeepSeek on raw price. Instead, it compresses its own pricing ladder. GPT-6.1 Sol is billed at one-fifth the standard input and output token prices. OpenAI says its performance approaches GPT-6 Astra on agentic coding, computer use and professional work 4. The company claims gains over GPT-6 Sol in debugging, document analysis, multi-step workflows and factual accuracy 4.

The headline reliability figure is a drop in error rate at low reasoning effort, from 11.4% to 7.7%. Across all reasoning settings, the model reportedly stays within 1.9% of Astra's error rate 4. OpenAI also highlights agent safety behavior. It says Sol more reliably flags broken search tools, follows explicit restrictions and avoids unauthorized outcomes 4. These are vendor-reported figures and should be checked independently. Still, they show OpenAI's argument. A model that is cheap but unreliable in agent loops is not actually cheap.

What it means

The two releases point in the same direction from different starting points. DeepSeek is competing on open weights, aggressive caching discounts and time-of-day pricing that rewards batch jobs run off-peak. OpenAI is betting that developers will pay its prices if a mid-tier model gets close enough to its flagship, with fewer of the failures that are expensive in production.

The practical takeaway for teams is that the per-token rate on a pricing page tells only part of the story. Cache-hit ratios, peak-hour exposure and agent error rates may matter more to the final bill than the advertised price. The fast arrival of third-party hosting for V4.1 Flash 5 adds another factor, since the same open weights may soon be available at prices that differ from DeepSeek's own rate card. With both labs lowering prices, the AI API price war has clearly moved from flagship models to the mid-tier models that handle most everyday coding work.