Three Flagships, One Price War
Three frontier model launches inside 48 hours. Anthropic's Fable 5 returned from export-control suspension on July 1 at $10/$50 per million tokens. Claude Sonnet 5 launched June 30 with intro pricing ($2/$10 through Aug 31, then $3/$15 standard). OpenAI's GPT-5.6 family emerged on July 9 with multiple cost tiers. The economics just shifted.
Three frontier launches in one window. Three different bets on how a frontier model should be priced. The leaderboard finally became a menu.
The three launches
Anthropic restored Claude Fable 5 on July 1, 2026 after the export-control suspension lifted. Pricing: $10 input / $50 output per million tokens, 1 million token context window. This is their frontier reasoning model, designed for tasks requiring deep inference.
Claude Sonnet 5 launched June 30 at introductory pricing — $2 input / $10 output per million tokens, valid through August 31, 2026. After Aug 31, pricing steps to $3/$15 standard. Sonnet 5 offers 1 million context and 128K output token limits, targeting the cost-efficiency bracket between Haiku and the frontier tier.
OpenAI's GPT-5.6 became the default ChatGPT model on July 9, rolling out three variants: Sol (highest capability, $5/$30), Terra (cost-efficient middle tier, $2.50/$15), and Luna (speed-optimized, $1/$6). All three share 1.05 million token context windows.
Why this matters
Frontier model competition is now pricing-driven. Anthropic's willingness to price Fable 5 at 5x Sonnet 5's output cost signals confidence in reasoning performance for high-ROI tasks. OpenAI's three-SKU rollout fragments the market — teams can now pick their tradeoff curve within a single provider. Google's Gemini 3.6 Flash (announced July 21 at $1.50/$7.50) entered this window at an even lower price point than Luna.
For teams: the days of picking one model and standardizing on it are ending. Economics now force portfolio strategies — Fable 5 for reasoning, Sonnet 5 for balance, GPT-5.6 Luna for speed, Gemini Flash for cost. The model choice isn't binary anymore; it's tactical.
The threat landscape expanded
Sysdig published analysis of JADEPUFFER, described as an agentic ransomware attack, showing that attackers holding only a DSN can take over a coding agent. The risk surface for deployed agents just widened — infrastructure secrets are now attack vectors to AI systems, not just backend services.
The Founder's Take
Three launches, three economics. Anthropic bet on premium reasoning pricing. OpenAI bet on portfolio segmentation. Google bet on aggressive volume pricing. The real competitive battle is no longer about capability — it's about who can own the cost curve and the use cases tied to it. If you're still running one model across every workload, you're leaving margin, latency, and reasoning quality on the table. Build the router before the bill forces you to.
— Megan
Get next week's note in your inbox.
🧠 Learn about AI from the frontlines. Follow us for AI updates in 10 minutes or less. Subscribe to Field Notes from the AI Frontier to stay ahead of AI news and updates. A founder's take as we consolidate the latest turning points and pivotal news from the top voices in AI. Learn how to build with AI and how to adapt.
🚀 Follow the future of AI with us.