Salud Capital Research · April 2025
As GPT-4 dominance breaks down and open-weight models reach parity on key benchmarks, the frontier is fragmenting. This is the current state of the model race, where each major lab stands, and what the pricing war means for anyone building on AI infrastructure.
Twelve months ago, GPT-4 was the unchallenged reference model. Every benchmark comparison started and ended there. That era is over. As of April 2025, the LLM market has fragmented into a genuine four-way race among OpenAI, Anthropic, Google DeepMind, and Meta — with a fifth disruptive force in the Chinese open-weight ecosystem threatening to commoditize inference entirely.
The fragmentation is not superficial. Each lab has identified a defensible niche: OpenAI on multimodal breadth and developer ecosystem lock-in through the API; Anthropic on enterprise safety and instruction-following reliability for long-context work; Google on infrastructure cost efficiency via custom TPUs and the Gemini 1.5 Pro’s 1M token context window; Meta on open-weight developer freedom through Llama 3. What’s collapsing is the idea that any single model can dominate all use cases simultaneously.
The price war is the structural story of the moment. GPT-4-class intelligence — which cost $0.12 per 1,000 tokens eighteen months ago — now costs approximately $0.01. Mistral AI offered a free API tier in early 2025. DeepSeek launched V2 at a price that undercut every Western provider by 60-80%. This deflation compresses the API-revenue moat and forces every lab to compete on ecosystem, not price.
| Provider | Flagship Model (Apr 2025) | Key Strength | Competitive Vulnerability | Price Point |
|---|---|---|---|---|
| OpenAI | GPT-4o (May 2024) | Multimodal breadth; developer ecosystem; ChatGPT consumer surface | Enterprise market share eroding to Anthropic; pricing pressure from below | $5/$15 per M tokens (input/output) |
| Anthropic | Claude 3.5 Sonnet (Jun 2024) | Enterprise reliability; long-context instruction-following; constitutional safety | No consumer surface; purely API-dependent; smaller ecosystem | $3/$15 per M tokens |
| Google DeepMind | Gemini 1.5 Pro (Feb 2024) | 1M token context window; TPU infrastructure cost; Workspace integration | Slow to ship; developer trust gap post-Bard errors; enterprise adoption lagging | $3.50/$10.50 per M tokens (<128K) |
| Meta AI | Llama 3 70B (Apr 2025) | Open weights; zero API cost; full customization freedom; massive install base | No managed inference product; requires self-hosting infrastructure | Free (self-hosted); $0.20-0.90 via third-party APIs |
| Mistral AI | Mixtral 8x22B / Mistral Large | MoE efficiency; European sovereignty angle; aggressive pricing | Smaller ecosystem; limited multimodal capability; funding vs. hyperscalers | $2/$6 per M tokens (Large) |
| DeepSeek | DeepSeek V2 (May 2024) | MoE architecture; 128K context; 60-80% cheaper than Western equivalents | China-based; data sovereignty concerns; US enterprise reluctance | $0.14/$0.28 per M tokens |
Meta’s release of Llama 3 in April 2025 in the 8B and 70B parameter sizes represents the most significant open-weight moment since the original Llama leak in February 2023. The difference is that this time it’s intentional, fully licensed for commercial use, and demonstrably competitive on real-world benchmarks rather than just leaderboard games.
Llama 3 70B performs comparably to GPT-3.5-Turbo on MMLU while requiring zero API cost for self-hosted deployments. Developers running on Akash, RunPod, or their own infrastructure can now access near-frontier capability without OpenAI or Anthropic taking a cut. For cost-sensitive applications — high-volume document processing, classification, code review — the economics shift materially.
The downstream consequence is price pressure on every closed provider’s sub-flagship tier. When Llama 3 70B can handle 80% of enterprise use cases at zero per-token cost, the GPT-3.5 and Claude Haiku markets face structural erosion. This forces closed labs up-market toward complex reasoning, multimodal, and long-context tasks where open-weight models still lag.
| Event | Expected Window | What It Resolves |
|---|---|---|
| GPT-5 / GPT-4.5 Launch | H2 2025 | Whether OpenAI can re-establish a meaningful capability gap over Claude 3.5 Sonnet and Gemini 1.5 Pro on complex reasoning |
| Claude 4 Family | H2 2025 | Anthropic’s next tier — whether Opus-4 can maintain enterprise coding leadership; whether Sonnet-4 can displace GPT-4o as the default workhorse |
| Llama 3.1 / 405B | Summer 2025 | Whether Meta can reach true frontier quality at open-weight; a 405B model competitive with GPT-4o would permanently flatten the closed-model pricing power |
| Gemini 2.0 / Flash | Late 2025 | Google’s cost-efficiency bet: if Flash delivers GPT-3.5-level at sub-$0.10/M pricing, it captures the high-volume enterprise tier |
| DeepSeek R1 / Reasoning | Late 2025 | Whether Chinese labs can deliver o1-comparable reasoning at open-weight, sub-$0.50/M cost — the event that would most disrupt Western API economics |