Cognition, the startup behind the AI coding agent Devin, is fielding offers that would value it between $70 billion and $80 billion, just weeks after closing a $2 billion round at a $48 billion valuation1,5. The same week, GitHub shipped Copilot Fusion, a model-orchestration layer that the company says matches the performance of Claude Code with Opus 5.5 Max at roughly half the API cost and 40 percent less time per task3,4. The AI coding market is splitting into two bets. One says a standalone agent company can command a valuation north of $70 billion on the strength of autonomous engineering workflows. The other says model orchestration is becoming infrastructure, and that the dominant developer platform can deliver it as a feature. The bets collide only if orchestration and autonomous agency compete for the same budget line; if they occupy different willingness-to-pay tiers, both can win. The valuation math, however, demands that Cognition's tier be very large.
Cognition's valuation trajectory is steep even by AI-era standards. The company raised $2 billion at a $48 billion valuation on September 8. Investors including Andreessen Horowitz, Accel, and Founders Fund participated. Cognition's existing backers also include Khosla Ventures, Elad Gil, and Pear VC. Now, according to people familiar with the matter, the startup is seriously considering a new round before year-end at a price between $70 billion and $80 billion. The talks are still early, and valuation targets could change. But the appetite is real enough that Business Insider's Katie Roof reported the company is actively weighing the offers. The comparison to SpaceX's $60 billion acquisition of Cursor reinforces the pattern: AI coding startups are being priced on trajectory and TAM conviction, not current cash flow.
ANALYSIS Without a disclosed revenue update since the mid-2025 run-rate figure in Vector Wire's tool profile, the gap between the last verified revenue number and the implied valuation multiple is itself a risk factor.
GitHub's Copilot Fusion, announced at GitHub Universe in San Francisco, attacks the problem from the opposite direction6. Rather than building a single frontier agent, GitHub built a routing and orchestration layer that selects among multiple models based on task complexity. Developers choose an effort tier (Standard, High, or Max), and Fusion handles which models run and how they coordinate. The system combines runtime model orchestration from Project HydraFusion, a research preview from September 2026, with Copilot Auto Model Selection.
The cost claim is the sharpest edge and the least proven. GitHub says Fusion Max delivered comparable benchmark performance to Claude Code with Opus 5.5 Max, scoring 60 on the Artificial Analysis Coding Agent Index (within two points of the Opus 5.5 Max baseline), at about half the average API cost per task and in about 40 percent less time. Crucially, this is a vendor evaluating its own product on a single index with no third-party replication. If Fusion's production performance diverges from the internal benchmark by even a modest margin on real-world task completion, the cost-parity narrative weakens: half the cost at lower quality is a different product, not a cheaper substitute. Independent verification of Opus 5.5's capabilities on other benchmarks shows the underlying model is strong: arxivlean_2026_09 scored 47.17 and brokenarxiv_overall scored 84.70. But those scores say nothing about how Fusion's routing preserves that strength across heterogeneous tasks. Jay Parikh, Executive Vice President of CoreAI at Microsoft, said Copilot Fusion coordinates multiple AI models to execute software development tasks, automatically selecting the most suitable approach based on the complexity and requirements of the work. GitHub is also restructuring its pricing: enterprise customers will pay for AI credits based on usage and will be able to access Copilot capabilities without purchasing a separate Copilot subscription.
The infrastructure commitment behind Fusion is substantial. GitHub is adding more than 4 million CPU cores and 160 petabytes of storage, following a previous expansion of 3 million cores and 120 petabytes in August after reliability incidents7. The platform recorded 7.38 billion commits in September alone, a fivefold increase over the previous month, driven by AI-generated code activity. GitHub recorded 34.3 billion code commits across 2026.
GitHub's distribution and its fresh infrastructure buildout represent a real advantage, but it is a distribution moat, not an orchestration moat. Orchestration is software that any well-resourced team can replicate. What GitHub can defend is the developer relationship and the data flywheel from billions of commits.
GitHub is not alone in treating model routing as infrastructure. Replit made its intelligent model routing system the default across every account over the summer. Ramp described moving model selection into Open SWE, its open-source coding agent, this month. The pattern is clear: model orchestration is becoming a commodity layer, not a differentiator. The value is migrating either upward, to the agent that owns the full autonomous workflow, or downward, to the platform that owns the developer relationship.
Mario Rodriguez, GitHub's chief product officer, described the company's goal as lowering the floor so more people can create software while raising the ceiling for professional developers. That framing explicitly positions Copilot as an augmentation tool, not a replacement. Cognition's pitch with Devin is closer to delegation: hand the agent a task and get back finished code. If enterprises want both, the market supports both.
But the cost dynamics complicate the coexistence narrative at current valuations. GitHub's half-cost claim, if it holds in production, puts direct pressure on the unit economics of any standalone agent that relies on frontier model inference. Cognition would need to demonstrate either superior task completion rates that justify a premium or a cost structure that can compete with a platform subsidizing orchestration across its entire user base.
Claude Opus 5.5, the model GitHub benchmarked Fusion against, carries list prices of $4 per million input tokens and $20 per million output tokens as of October 28, 2026. If Fusion Max uses roughly half the tokens per task while matching Opus 5.5 Max performance, the effective cost advantage compounds: fewer tokens at the same per-token price, or comparable tokens routed to cheaper models where frontier capability is unnecessary. That arithmetic makes orchestration a cost threat to any company whose margin depends on selling frontier inference, but only if the quality holds. Routing to cheaper models requires knowing in advance which subtasks are easy, a classification problem that itself introduces error.
Three things to watch. First, whether Cognition closes a round before year-end and at what valuation; the gap between $48 billion and $70–80 billion in under four months would set a striking velocity marker for private AI companies. Second, whether GitHub publishes independently verified benchmark results for Copilot Fusion; the current performance claims rest on GitHub's internal evaluation and a single index score. Third, whether Cognition discloses updated revenue or usage figures that could anchor the valuation to something more concrete than the mid-2025 run rate. Orchestration is commoditizing. The open question is whether the autonomous agent layer above it is a durable product category or a feature that platforms absorb next.