Claude Opus 5.5 and GPT-6 Sol: The AI Price War Explained
Yesterday afternoon, the two most important AI companies in the West got into a public price fight — and small businesses are the quiet winners.
At 4pm on September 22, Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens — roughly 40% cheaper than Opus 5 on typical workloads, with output generation 30% faster. It matches Claude Fable 5.1 on most tasks, posts 66.4% on Terminal-Bench 4.0 (up from Opus 5’s 52.3%), and hits 81.8% on OSWorld 2.0 computer use.
Roughly 90 minutes later, OpenAI answered with two new models: GPT-6 Sol and GPT-6 Luna. Sol is priced at $2/$10 per million tokens — half the previous GPT-5.6 tier — aimed at complex coding, with OpenAI claiming about half the mistakes of its predecessor on internal factuality evals. Luna targets high-volume clerical work at the same cut-rate pricing.
Two frontier labs. Two flagship-tier releases. One evening. Both at the lowest price-per-capability we’ve ever seen from them.
Why this matters if you run a small business
You don’t buy tokens. So why should you care about a pricing war between API vendors?
Because token prices are the wholesale cost of every AI product you’ll ever buy. Every receptionist agent, inbox triage tool, and follow-up automation is priced on top of model inference. When wholesale drops 40–50% overnight, retail follows — cheaper plans, higher usage limits, or both. The cheap-model revolution we tracked when DeepSeek V4-Flash launched just arrived at the top of the market, not just the bottom.
It also means “always-on” AI stops being a budget line you have to defend. A year ago, running a capable agent 24/7 on a frontier model was expensive enough that you’d think twice. At half the price — with models that make fewer mistakes and need less human babysitting — the math flips: the hours saved cost less than the tokens burned.
The strategy behind the cuts
The price drops aren’t generosity. Three forces are colliding:
Open-weight competition. Free, self-hostable models keep closing the gap — Arcee AI hit near-frontier quality at 96% lower training cost, and hundreds of companies now back open-weight AI. Frontier labs can no longer defend premium pricing on capability alone.
Collapsing customer margins. Bloomberg reported this week that Harvey, Abridge, and Ramp — major AI application companies — have started ditching frontier-lab APIs for open-weight models after their margins collapsed to negative 50%. When your biggest customers have a credible exit, you cut prices to keep them.
A land grab for agents. The next battleground isn’t chatbots — it’s autonomous agents doing real work, which consume 10–100x more tokens than a conversation. Whoever prices lowest becomes the default infrastructure for that wave.
What doesn’t change
Cheaper doesn’t mean free, and it definitely doesn’t mean automatic ROI. The failure mode we see most often is unchanged: businesses that buy “AI” as a strategy instead of pointing it at a specific, recurring, measurable task.
The price war helps you whether you use it or not — your vendors’ costs just fell. But the businesses reporting real returns are still the ones that deployed agents against missed calls, inbox overload, and follow-up drudgery, not the ones waiting for prices to fall further.
Prices will keep falling. Waiting for the bottom is a strategy that never triggers.
The bottom line
Frontier AI capability at half price, twice in one evening, is not a flash sale — it’s the new trajectory. Every AI-powered tool your business touches will get cheaper, faster, and more capable on this curve. The advantage now goes to businesses that have already wired AI into their operations and simply ride each price cut down, versus those still deciding.
If you’d rather ride the curve than watch it, SquidBot builds and runs AI agents for small businesses — receptionists, inbox triage, and follow-up — and we pass every model price cut straight through to the automations we run for you.
