For four years, the frontier-model story was simple: the lab that scales fastest wins. OpenAI trained the largest models. Anthropic and Google trained smaller models. The performance gap justified OpenAI’s strategy and market valuation. That era ended in April 2026. On April 24, OpenAI released GPT-5.5; on February 20, Google released Gemini 3.1 Pro; on April 16, Anthropic released Claude Opus 4.7. All three now converge on equivalent frontier performance. The race is over. The moat has collapsed. And the labs that spent the least compute to reach the frontier now have better unit economics than the lab that spent the most.

This is the most consequential inflection in AI markets since the release of GPT-3.5. It does not mean OpenAI is failing. It means the competitive game has shifted, winner-take-most dynamics are reversing, and the next 18 months will determine which labs survive the transition to an efficiency-driven market.

The Capability Convergence Is Real

The evidence is specific. On major benchmarks, the three models now cluster within 1-3 percentage points of each other. Opus 4.7 scored 64.3% on SWE-bench Pro (a real-world software engineering benchmark), versus 58.6% for GPT-5.5, giving Anthropic the coding advantage. GPT-5.5 scored 82.7% on Terminal-Bench 2.0 (a test of command-line and system tasks), versus 69% for Opus and 68.5% for Gemini. On scientific reasoning (GPQA Diamond), Gemini 3.1 Pro tied Opus 4.7 at 94.3%, just ahead of GPT-5.5 at roughly 93-94%. No model dominates every category. Each wins in a different lane, and the leads are measured in single digits.

On the Artificial Analysis Intelligence Index, an aggregate benchmark combining multiple domains, GPT-5.5 scores 60, while Claude Opus 4.7 and Gemini 3.1 Pro both score 57. A three-point gap at this level of performance is within margin of error for real-world workloads, where task-specific tuning matters more than aggregate scores.

This convergence should have been impossible if the scaling hypothesis was correct. The hypothesis held that training compute directly predicts capability, and labs with more compute would have permanent edges. Anthropic trained Claude Opus 4.7 on substantially less compute than OpenAI used for GPT-5.5 (estimated 3-5x variance). Yet Anthropic achieved equivalent results. That violates the moat story. It proves the moat story was always incomplete.

The Real Constraint Is Now Economics, Not Capability

What changed is the shift from training-time competitive advantage to inference-time competitive advantage. Training a frontier model is now a solved problem for any lab with $10+ billion in capex. The constraint was never the research; it was access to capital and compute. Four labs have that: OpenAI (via Microsoft), Anthropic (via Google and Amazon), Google (internal), and Meta (internal). Chinese labs (Alibaba, ByteDance, Baidu) have access via state subsidy.

But building frontier capability is 10% of the problem. The other 90% is inference cost, latency, and ecosystem stickiness. OpenAI charges $5 input/$30 output per million tokens; Anthropic charges $5 input/$25 output; Google charges $2 input/$12 output. At parity capability, a customer serving millions of API calls will migrate to the cheapest option (Google or Anthropic) unless OpenAI’s ecosystem (ChatGPT brand, plugin ecosystem, enterprise distribution) is worth the 2-3x price premium. For many enterprises, it isn’t.

The lower-cost labs also benefit from improved unit economics on marginal customers. OpenAI spent $150+ billion on capex over three years to achieve frontier capability. Spreading that cost across inference customers means OpenAI needs massive query volume to break even. Anthropic, by contrast, outsourced training compute to Google’s capex, meaning Anthropic’s marginal cost on each inference query is lower. That allows Anthropic to undercut OpenAI on price and still hit higher gross margins.

Challenger Labs Are Already Winning on Efficiency

This opens a 12-18 month window for labs that optimize for inference efficiency rather than frontier capability. DeepSeek’s V4 Flash model achieves near-frontier performance at a fraction of the cost: $0.14 per million input tokens and $0.28 per million output tokens, roughly 15-20x cheaper than GPT-5.5. That pricing structure reflects substantial capital efficiency gains, whether from state subsidy or algorithmic innovation. DeepSeek doesn’t need frontier market share. It just needs to be “good enough” for customers who value cost over the last 5% of capability.

Mistral, with similar efficiency focus, is positioned similarly. Both labs are competing for the long tail of AI workloads: customer support, data labeling, synthetic data generation, routine inference tasks that don’t demand frontier reasoning. These workloads represent 80%+ of total inference queries but generate only 20% of API revenue for frontier labs. Challenger labs can build sustainable, profitable businesses by owning that segment.

The Historical Parallel: Intel vs. AMD

The closest historical parallel is Intel’s dominance in the 1990s CPU market. Intel owned the architectural advantage and manufacturing lead. AMD was perpetually 2-3 years behind. Then AMD secured a partnership with TSMC (Taiwan’s leading chip manufacturer) and modern foundry access. The technology gap didn’t close overnight, but it closed enough. AMD achieved near-equivalent performance, and competition shifted from “whose chips are faster” to “whose chips are cheaper and more reliably available.” Intel’s gross margin collapsed from 65% to the low-40s. Intel remained the market leader, but its share fell from 80%+ to 55-60% as AMD and others captured the rest.

The dynamics apply directly to AI. OpenAI’s “first mover” advantage in frontier capability (analogous to Intel’s architectural lead) is now neutralized by capability convergence. If Anthropic (subsidized by Google’s compute), or DeepSeek (subsidized by Chinese capital), or any other lab can train frontier capability at 1/3 the cost, the margin story inverts. OpenAI no longer has the luxury of high gross margins on inference. It must compete on price, ecosystem, and brand. That’s a weaker position than dominance through raw capability.

What Happens to Market Structure

Expect the frontier-LLM market to consolidate into an oligopoly within 18 months. The four global labs with $10+ billion in capex (OpenAI, Anthropic, Google, Meta) will persist. Chinese labs will capture 15-25% of the market via lower cost. Smaller labs (Mistral, xAI, Stability, others) will focus on specific verticals (code, multimodal, reasoning) where efficiency specialization creates defensibility. The “winner-take-most” narrative that dominated 2023-2025 is obsolete. The market is fracturing into segments, each with its own leader based on cost curve, distribution, and use-case fit.

Hyperscalers will face margin pressure. Gartner estimates that by 2030, inference cost will fall 90% from 2025 levels. That’s an existential threat to anyone betting on inference-API revenue as the primary business model. Labs that can shift revenue toward vertical-specific applications (customer service, code generation, research automation) will sustain margins. Labs that try to compete on raw inference price will lose.

The April 2026 releases marked a regime change. For the first time since GPT-3.5, the frontier-model market has shifted from a race (who scales fastest) to a competition (who has the best unit economics). Winners will be determined not by research prowess but by capital efficiency, distribution access, and ability to sell outcomes, not tokens.