4 AI Coding Agents Share One Sandbox-Escape Flaw, Patch Now

A customer asks a voice bot a question and waits. At around 2.5 seconds of silence, the bot stops sounding thoughtful and starts sounding broken. The threshold is physiological, not technical, and models that cannot hit it are out of the running for this workload. NiCE, which runs deployments for Toyota, Allianz, Nestlé, and Frontier, clocked Anthropic's Opus 4.7 at over 4 seconds to first token on real customer traffic. Gemini 3 Flash comes in near 1 second, at roughly 1/18.5th the cost.

The complication that keeps this from being a Google-wins story: Anthropic holds roughly half the error rate of Gemini 3 Flash. What is actually on the table is an accuracy-versus-latency tradeoff, and the correct product answer is tiered routing rather than a vendor pick. Cheap, fast models take the high-volume, low-stakes turns. Accuracy-optimized models get reserved for escalations and compliance-sensitive replies, where a wrong answer costs more than a slow one.

Google is pressing on the fast tier. The new Gemini 3.5 Flash-Lite hits 350 output tokens/sec as the cheapest option in its lineup, built specifically for latency-sensitive agent loops, and 3.6 Flash cuts output tokens by about 17%. OpenAI is in the same race from a different angle: its cheaper mini models are already gaining share among NiCE's customers, and its bidirectional, interrupt-capable audio model, still ChatGPT-only, could reset the comparison once it reaches the API.

The practical shift for anyone shipping voice or chat features: 2.5 seconds is now a line in the PRD, not a benchmark footnote. Any feature that inherited a vendor's default flagship model should be treated as out of spec until a latency test on real traffic says otherwise. The useful move this sprint is to run that test before the next release rather than after the first complaint, and to decide in advance which conversation turns are allowed to be slow.

The engineer scanning Arena's frontend-coding board this week sees Kimi K3 on top, priced at half the per-token rate of GPT-5.6-Sol. What the board does not show is that K3 burns roughly 2x the tokens per task, which puts net cost close to flat. It also ships with a single 'max effort' reasoning mode, and that mode spent 13,000+ tokens drawing one icon. A tester's verdict covers it: 'Incredible model, not an incredible value.' The unit that matters is cost per completed task, and on that unit the leaderboard win evaporates.

Orchestration is the larger lever. Cursor's SQLite rebuild produced an 8x cost spread ($1,339 vs $10,565) across configurations that all passed identical tests, which means the setup was the variable, not the model. Ramp's internal router, an EWMA failure-rate tracker paired with Thompson-sampling latency selection, already has a documented 30% cost reduction with no performance loss. No model swap on offer this quarter moves the number that far.

Pricing structure is moving underneath all of this. Sierra's Horizon now charges by business outcome achieved rather than tokens consumed, and that model will surface as a sales objection within a year, phrased roughly as 'why pay when it doesn't work?' A quieter fix lands next week: MCP replaces client-issued session IDs with a stateless, server-side model, closing the scaling bug that breaks load-balanced agent deployments.

Teams still debating which model to wrap are optimizing the smallest variable in the stack. Unit economics move at the routing layer and in the completion-cost math, with the pricing model as the next front. The exercise worth running before the next sprint is to divide last month's total spend by completed tasks, then produce the same figure under a second routing configuration. That comparison decides more than any leaderboard position does.