# Mozilla Open Source AI Report: The 28-Point Deployment Gap > Source: > Published: 2026-07-18 10:10:13+00:00 Mozilla dropped its inaugural [State of Open Source AI report](https://stateofopensource.ai/) this week, and the headline number reframes everything: 79% of developers already use open models, but only 51% have shipped one to production. The performance gap between open and closed AI is 3.3%. The deployment gap is 28 points. So the question isn’t which AI is better anymore — it’s why open models keep dying between the demo and the deploy. ## The Model War Is Basically Over The capability gap between open and closed models has collapsed. GPT-4-equivalent inference fell from $20 to $0.40 per million tokens in three years — a 50x reduction. Open models now match closed systems in coding, instruction-following, and general knowledge. The one area where closed still wins is long-context retrieval: Gemini 3 hits 89% multi-needle accuracy at 1 million tokens; DeepSeek V4-Pro lands at 41%. If your application reads 200-page documents end-to-end, that gap still matters. For most workloads, it does not. The economics have been settled for longer than the industry admits. 37signals moved off expensive cloud AI and cut its bill from $3.2 million to under $1 million. GEICO’s AI spend ran 2.5x over projections before the company restructured its contracts. Uber engineers burned through $500–$2,000 per month in AI coding tool budgets until the company capped usage at $1,500 per month. The Linux Foundation estimates $24.8 billion in annual unrealized savings if enterprises shifted suitable workloads to open models. ## So Why Aren’t Developers Shipping Open Models? The [report surveyed 950+ developers globally](https://blog.mozilla.org/en/mozilla/mozilla-state-of-open-source-ai-report/) and found five operational blockers dominating every region: infrastructure costs (27%), security and compliance (26%), maintenance burden (24%), deployment complexity (23%), and a shortage of specialist support (22%). None of these are model quality problems. The most telling data point is the enterprise scaling asymmetry. Closed model production success climbs from 54% at small companies to 73% at organizations with 1,000+ employees. Open model success barely moves: 53% to 57% across the same scale. Large companies can buy their way through closed-model deployment — vendor support, SLAs, dedicated implementation teams. Open model deployment waits on tooling that nobody has finished. Developers with vendor support achieve a 67% production success rate with open models. In-house only: 33%. That isn’t a model problem. It’s an infrastructure problem wearing a model’s clothing. ## The Agentic Layer Is Where It Gets Complicated The report’s most forward-looking finding: the model matters less than the harness around it. Mozilla defines the “agentic harness” as the software controlling what an AI system can access and execute. In third-party benchmarks, open scaffolds outperformed proprietary harnesses — Terminal-Bench 2.0 showed third-party setups at 79.8% vs. 58.0% on the same underlying model. Labs have since closed that gap by optimizing models for their own harnesses, while keeping the 6x price premium firmly in place. Closed platforms aren’t just selling models. They’re welding model and scaffold into one rented product. Once your production agent stack depends on a proprietary harness, switching costs go from annoying to architectural. Meanwhile, the “write surface” — the set of irreversible actions an AI agent can take — has no portable standard. No agreed specification governs when agents may transfer money, modify records, or delete files across framework boundaries. The result: users approve AI agent requests 93% of the time by default, a phenomenon the report calls consent fatigue. The [Model Context Protocol](https://spec.modelcontextprotocol.io/) has 97 million monthly SDK downloads and 10,000+ active servers. Governance maturity across companies using it: 21%. The infrastructure is growing faster than the guardrails. Mozilla’s response is not to build another base model. The organization is building the harness layer itself — software to enforce policy, approval thresholds, and cost caps above the model, independent of which model sits below it. ## What Developers Should Actually Do Three moves worth making now. First, run a [total cost of ownership calculation](https://stateofopensource.ai/state-of-open-source-ai-2026.pdf) before your next renewal — per-token pricing is visible, but vendor lock-in, harness dependency, and data residency constraints accumulate quietly. Second, audit your agent governance before you scale. If your system approves 93% of agent actions by default, that’s not permission — it’s automation theater. Define your write surface explicitly, even informally. Third, watch the harness consolidation. The window to build on genuinely open infrastructure before closed platforms lock the scaffold layer is real, and it is not permanent. Open models are almost there on capability. Deployment infrastructure is not. That gap is where the next year of AI development will be decided — and it won’t be the model vendors who determine the outcome.