August 2026 Summaries

Data from the Weave model router shows that fixed model choices for AI coding agents are frequently inefficient: across roughly 275,000 requests from 22 installations over 30 days, 79.2% were routed to a different model than originally requested. The router commonly directs routine tasks such as classification, title generation, and tool-output summarization to cheaper, faster models, while also downshifting some frontier-model requests when their task requirements do not justify higher capability. These decisions reduced routed-traffic costs by about 28% without requiring configuration changes or reportedly causing quality regressions, though routing added a median 329 milliseconds, or roughly 7%, to a typical 4.8-second upstream request. Users explicitly forced a model in 12.4% of requests, an override mechanism presented as necessary for retaining control when specific model preferences matter.

For coding-agent workloads, LLM routing should account for session-level prompt caching rather than treating each request independently, because 62.1% of the 46 billion input tokens handled by the Weave router over 30 days were discounted cache reads. Since provider caches are tied to a model and an exact conversation prefix, moving a session to another model requires resending its accumulated context at full input cost, often outweighing apparent per-request savings. Weave therefore makes a substantial routing choice on the first turn and pins later turns to that model by default, with 14,316 pinned sessions averaging 10.4 turns, 33.6% of requests served through pins, and only 1.4% of pinned sessions switching models. Low failover rates and frequent agreement between pinned models and shadow routing choices suggest that most initial selections remain appropriate, although switching can still be worthwhile for provider failures, repeated agent errors, or a shift to more complex work. The approach argues that chat traffic may suit request-by-request routing, while agent systems need to price cache loss explicitly and use deliberate session un-pinning decisions.

Claude Opus 5 was added to the Weave router in late July and grew from eight requests in its first week to more than 12,000 weekly requests within five weeks, reaching 10.8% of router traffic without a fleetwide manual rollout. The router uses a bandit policy that classifies requests by task type and routes them among competing models based on measured success, cost, speed, and failure patterns, allowing new models to begin with cautious traffic informed by offline evaluation and gain usage where live results justify their cost. Safety controls restrict experimentation to suitable fresh sessions, while user-selected models, designated low-cost internal tasks, and ongoing sessions remain pinned. Routing data indicated that Opus 5 became concentrated in longer, complex multi-file sessions rather than broadly distributed across users, while lower-cost models such as deepseek-v4-flash handled subagent tasks like codebase exploration. The approach frames frontier-model adoption as a continuously measured routing decision, enabling models to earn or lose traffic gradually with transparent records of each choice.

Weave’s router analyzed coding-agent activity in shadow mode over 30 days, monitoring roughly 275,000 requests across 25,500 sessions for behavioral failure patterns such as repeatedly editing the same file or continuing after consecutive tool errors. It recorded 582 spiral events and 161 broader struggle signals, indicating that such failures are uncommon but meaningful at scale. Same-file thrashing appeared across all monitored models, typically involving five to seven edits to one file, suggesting that looping is a general agent-loop failure mode rather than a weakness unique to a specific model; error streaks varied more by workload, with DeepSeek V4 Flash recording the most. Strict byte-identical tool-call loops were very rare, with only a few cases, and routing those sessions to a stronger model reportedly resolved them quickly. Meanwhile, malformed tool calls were nearly absent across more than 150,000 tool-use blocks, with the highest invalid-argument rate at 0.04%, leading the analysis to argue that agent reliability concerns have shifted from tool-call syntax toward detecting semantically unproductive behavior over time.

Over a 30-day period, Weave routed about 235,000 requests across roughly 40 AI models using complexity clusters—fast, balanced, high, and maximum—alongside hard-pinned utility tasks, user-selected models, and legacy routing rules. DeepSeek-V4-Flash handled the largest share of requests by serving inexpensive fast tasks, subagent bootstraps, and tool-result follow-ups, while Gemini 3.1 Flash Lite was used exclusively for pinned classification, title generation, and health checks. Minimax-M3 dominated routine balanced tasks under a newer XGBoost-based selector, whereas Claude Sonnet 5 led high-complexity coding work and consumed the most tokens. The maximum tier, including Claude Opus 5, Claude Fable 5, GPT-5.5, and GPT-5.6-Sol, handled fewer but more costly and often longer-running sessions, with premium models retaining users across many turns. The data shows a large cost imbalance: the cheapest models processed about 38% of requests for roughly 1% of spending, while Claude models represented about 89% of total spend despite handling about 38% of requests, partly offset by higher cache-read rates. Model usage changes quickly as models are introduced, replaced, or gain learned routing preference, illustrating that the router continuously adapts rather than relying on fixed assignments.

Weave optimizes the routing of requests to machine learning models by selecting the most cost-effective model that can handle each specific task, thereby substantially reducing costs while maintaining performance. By routing entire sessions rather than individual requests, Weave preserves the working cache, ensuring that cost savings are real rather than theoretical. This approach results in significant cost reductions, such as an 88.5% cut on median requests and a 68.7% cut across all routed-down requests, while maintaining reliability with a 98.8% success rate for failover requests. Unlike naive routing practices that switch models to save money but end up increasing costs due to cache loss, Weave's session-based routing ensures that these savings are sustainable in production environments. The system is designed to quickly adapt to new model releases without requiring code changes, allowing users to benefit from improved models and cost savings seamlessly.