OpenAI Just Made Its Flagship Cheaper Than Anthropic — And It's a Trojan Horse

The real target isn’t developers. It’s your CFO.

OpenAI announced on August 22 that it is cutting the price of GPT-5.6 Sol, its flagship model, by at least 20% for the next three months. The discount applies only to API and credit-based usage — not to ChatGPT Plus, Pro, or Business subscribers. That distinction matters. This is a B2B play, and it is aimed squarely at enterprises that bill AI costs back to engineering budgets.

The new rate: $4 per million tokens for input, $20 per million for output. For comparison, Anthropic’s Opus 5 sits at $5 and $25 respectively. Even before the cut, Sol was competitive. After it, OpenAI’s top model is now the cheapest flagship among the major players — and that is not an accident.

What makes this cut sharper than it appears is the volume structure underneath. OpenAI has also expanded tiered discounts for high-usage accounts, meaning enterprises running large-scale workloads will see effective rates even lower than the headline numbers. A company processing ten billion tokens monthly could pay closer to $2.80 per million input tokens — a margin that leaves virtually no room for rivals to match without writing off their infrastructure costs.

Three moves in one week

This is the third pricing strike from OpenAI in eleven days. On August 30, it cut GPT-5.6 Luna by 80% and Terra by 20%. At the time, observers noted the pattern — OpenAI was pruning its own product line to remove price gaps that gave Anthropic an opening. Today’s Sol cut closes the final gap.

The sequence tells a clear story. First, commoditize the mid-tier so competitors cannot gain traction on price. Second, make the flagship cheaper than the competition’s flagship. Third, reward the users who are actually driving volume.

Breaking it down further: the Luna 80% cut eliminated the most natural upgrade path for mid-tier customers. Someone evaluating Luna versus Opus 5 found the math suddenly indifferent. The Terra 20% cut then collapsed the gap between mid-range and high-end performance per dollar. By Wednesday, there was nowhere left to hide.

Third-party data already backs the strategy. On OpenRouter, a multi-provider AI hosting platform, GPT-5.6 model usage has roughly doubled Anthropic’s latest offerings, even as OpenRouter itself is discounting GPT-5.6 prices by half. Anthropic’s flagships, burdened by higher per-token costs, have struggled to keep pace. The flywheel is visible in the routing logs — every dollar of OpenAI’s price reduction is translating directly into shifted request volume.

The real weapon: Codex

Price cuts alone would be a defensive move. What makes this offensive is what OpenAI is pairing it with.

Codex, the AI coding agent, crossed 20 million weekly active users on August 21, according to project lead Tibo Sottiaux. That is up from 15 million just three weeks earlier — a gain of 5 million users in roughly 20 days. To celebrate, OpenAI is crediting every Codex and ChatGPT Work user with a “banked reset” they can use at their discretion.

The timing is deliberate. Codex is the consumption vehicle. GPT-5.6 Sol is the pricing lever. Together, they create a flywheel: lower API costs pull in enterprise developers, Codex turns those developers into daily power users, and daily usage justifies larger contracts — all while lock-in increases.

Consider the second-order effect: every Codex session generates code suggestions, debugging help, and architecture recommendations. That usage trains OpenAI’s internal feedback loops and generates proprietary telemetry about how developers actually work. Anthropic gets raw tokens. OpenAI gets a behavioral map of the global programming workforce.

This is how cloud providers won the infrastructure war in the 2010s. Undercut on compute, over-deliver on tooling, and let usage compound. OpenAI is playing the same game, except the commodity is intelligence instead of processors — and the switching costs are measured in retrained pipelines, not migrated databases.

Who loses

Anthropic is the obvious answer, but the damage may be broader.

Companies that signed enterprise AI deals at last quarter’s prices are now looking at contracts that overpay relative to the market. OpenAI is effectively repricing the category mid-cycle, and no competitor can match a 20% cut on a flagship without eroding their own margins. Those that hold price risk losing enterprise mindshare to OpenAI’s bundled value proposition — model + tooling + usage momentum.

The ripple effects extend beyond direct competitors. Integration partners and system integrators who recommended Anthropic-based architectures six weeks ago now face awkward conversations with clients who see OpenAI’s new pricing. Custom fine-tunes, RAG pipelines, and evaluation frameworks built around Opus 5 are suddenly anchored to a model that looks expensive by comparison — even if the quality differential justifies it, justification is harder to sell than savings.

Smaller AI startups face an even steeper problem. They cannot compete on price at this tier, and they cannot match the distribution of a platform with 20 million weekly coding agents. The bar for differentiation keeps rising. The question hanging over every Series A AI startup is whether their technical edge survives when the baseline becomes this cheap.

The sub2api loophole

Not everything is clean. OpenAI acknowledged that some users are exhausting their usage limits prematurely, and attributed much of it to a pattern the company calls “sub2api” — converting subscription access into API calls, which carry different limits and higher throughput. OpenAI described the issue as a system optimization rather than a bug, and pushed back against community claims of a breakdown.

This is a familiar tension in platform economics. Free or discounted entry points attract power users who find ways to extract more value than intended. The question is whether OpenAI will close the loophole or let it ride as a customer-acquisition tool. So far, the company appears to be doing both — publicly acknowledging the pattern while quietly benefiting from the attention it generates.

From a product perspective, sub2api reveals something important about demand elasticity. Users are already pushing hard against usage caps at current pricing. Make the API cheaper, and that pressure intensifies. OpenAI may be betting that the conversion funnel from loose API users to paid enterprise contracts is wide enough to absorb the leakage — that every sub2api power user is a future logo on an enterprise deal.

What happens next

OpenAI’s three-month discount is temporary, but the market shift it creates will not be. Once enterprises re-architect their pipelines around GPT-5.6 Sol’s pricing, switching costs rise. Once Codex becomes a daily habit for 20 million developers, the platform stickiness compounds. The discount expires in November. The habits don’t.

Anthropic will need to respond — either with its own cuts, a repositioning around reliability or safety, or a push into verticals where OpenAI’s price advantage matters less. We are already seeing early signals: Anthropic has been quietly expanding its enterprise support tiers and emphasizing Claude’s longer context windows and more deterministic outputs. But those differentiators only hold weight if buyers believe they matter more than a 20% cost gap — and right now, CFOs are doing the weighing, not CTOs.

Google, Microsoft, and Amazon will watch closely and likely adjust their own positioning. Google’s Gemini and Microsoft’s GPT-integrated Stack are already priced below OpenAI’s pre-cut rates, but they lack the Codex distribution engine. Amazon’s Bedrock multi-model approach gains new urgency when one model category suddenly drops 20% overnight — customers will rebalance their routing, and Bedrock’s leverage depends on keeping all providers competitive.

The broader implication is structural. The enterprise AI market is moving from a scarcity model — where capability gaps justified premium pricing — to a surplus model where the differentiator is distribution and integration depth. OpenAI is betting that whoever controls the developer workflow controls the pricing floor. If Codex hits 30 million users by year’s end, that bet looks less aggressive and more inevitable.

The arms race is no longer about who builds the smartest model. It is about who can afford to sell it the cheapest — and still make money on the way up.