Coding models get faster, cheaper, and more spread out

A faster model at a lower price

Cognition's team introduced SWE-2 Flash, a coding model built with a new mid-training stage before reinforcement learning. It is available through Devin Desktop and the command line, and the company says it is one-fifth the price of SWE-2 and three times as fast. Those two numbers together are the pitch: a model cheap enough to run everywhere and fast enough to keep an agent loop moving. m:…5916

The company reports large benchmark jumps over its predecessor, Inkling-Small: Terminal-Bench 4.0 from 0.9% to 20.0%, FrontierCode 2.0 from 3.2% to 24.0%, and an ExploitBench score up from 11.3% to 34.8%. Benchmarks are vendor-reported here, and they measure the model on tasks the company chose to highlight, so they are a signal, not a verdict. Still, the direction is consistent with the price and speed claims: this is a model meant to be used constantly, not sparingly. m:…5916

Cognition is building a program to share weights with industry partners and researchers, but says it is not sharing them publicly because of cyber risks. That split matters for buyers: API access now, partner weight access later, and no public weights to self-host in the meantime. m:…5916

Inference that tunes itself

Together AI's answer to the same cost problem is to stop hand-tuning inference at all. Its Together Compiler is an agentic system that continuously adapts inference workloads to changing traffic, models, and hardware, using persistent engines that start once and hot-reload candidate forward passes instead of restarting. A two-tiered design uses fast engines for rapid iteration and slower engines for high-fidelity testing. m:…0495

The company says the compiler reaches up to 2.6x higher tokens per second across frontier open-source models, and 252x higher experiment throughput than baseline agentic optimization systems. It also notes that users' prompt lengths grow by 80% two weeks after a new model release as they shift toward agentic workloads, which is exactly why a static configuration goes stale fast. m:…0495

Chutes, meanwhile, expanded its catalog with seven new models from Z.ai, DeepSeek, Alibaba, and Moonshot AI, all reachable through its existing OpenAI-compatible API. The cheapest listed entry is DeepSeek V4.1 Flash 1017 at $0.30 per million regular input tokens, and Kimi K2.7 Code is reported to use about 30% fewer thinking tokens than Kimi K2.6. The catalog, not any single model, is the product: one API, many price points. m:…8097

Developers already route around loyalty

OpenRouter data cited by Kilo Code shows how developers actually behave. Customers spending $100 or more a month now use a median of 6 model families, double the figure from a year ago, and the median across customers climbed to 7.3 in October 2026 after bottoming at 3.8 in June 2025. Customers spending $10,000 or more a month use a median of 12 labs. The pattern is deliberate: cheaper models write and fix code, and a stronger model reviews it. m:…1407

Revolut is the extreme version of that pattern. Its self-hosting bill is one-tenth of what it spent on Anthropic and OpenAI, it measured a 70% improvement in lead time after moving to self-hosted GLM 5.1, and it now serves a self-hosted open model, GLM 5.3, as the default coding model for 2,000 engineers on 56 Nvidia B200 GPUs. Self-hosting is not free, but at that scale the company says it costs about 30 cents per million tokens. m:…1407

The takeaway for a team setting a coding budget is to stop asking which single model to buy. No single model takes more than 30% of any agent role usage within Kilo, and Nvidia's Nemotron models lead in code, plan, and debug tasks there. The portfolio is the strategy, and the routing between models is where the money is saved or wasted. m:…1407