Talon optimises the loop the agent runs in. It predicts what the task will touch, serves the right context up front, and steers clear of the loops that burn turns: same model, same prompts, measurably less spend. Live now, measured under a sealed benchmark.
Every figure on this page traces to a sealed benchmark cell: both arms best-configured, workspace sealed, treatment verified on the transcript, quality judged by held-out tests. We publish where we lose, too. Full methodology · per-task breakdown →
We measured what coding agents do with their turns, on real work in real repositories. Three patterns, over and over.
Files re-read, searches re-run, ground re-covered. Across 9,459 real agent traces, 47.5% of the steps served nothing the task needed. That is spend with no output attached.
The agent maps the same code it mapped yesterday, then re-checks what it had already established. Two loops. Both burn turns you pay for, and neither writes a line of code.
On a large codebase the agent takes 14 tool calls where 4 would do. Grep doesn't recognise structure, so the agent walks the repository the long way, and the meter runs the whole time.
Talon puts the right context in front of the model before it starts, so the loops above never open. It predicts what each task needs, serves it up front, and steers the agent clear of the turns that burn budget. A prediction and heuristics engine, not just a graph.
Same model. Same prompts. Same result. Measurably less spend — and 2× fewer broken tests than the baseline while doing it.
Get alpha access →Talon keeps a structural map of your code, current as you commit, and puts it to work in three moves.
The structural context a change actually touches — what flat search and embeddings miss. Predicted per task.
Served before the agent's first turn, so it starts informed instead of exploring. No discovery phase.
Steered away from the rediscovery and re-verification loops that burn turns. Every intervention logged, so you can audit what it did.
A small thin client installs locally and changes one thing: the target your harness sends its traffic to. From there your harness does the routing — straight to the Cogmeta proxy, which forwards it to Anthropic. Headers forwarded, not forged. We don't route your traffic; we just change where it's pointed. Nothing else about your workflow changes: same agent, same models, same prompts.
Because being in the request path is what unlocks the stack. In the loop, Talon can put the right context in front of the model before a turn starts, steer the agent while it runs, and measure every run end to end. Out of the loop, none of that is possible.
One install, and a hop through our proxy on the way to Anthropic. That's it. Your agent, your prompts and your models are untouched, and every claim we make about the result is measured on the same traffic.
They build context into their own editor or agent — you switch to their tool to get it. Talon is a layer for the agent you already run: Claude Code today, any harness next. It optimises the whole loop for cost, which a bundled editor doesn't, and it charges nothing for the tokens it saves you. Keep your tools. Talon just makes them cheaper. Complementary, not competing.
Free during the private alpha — early users grandfathered. Usage-based pricing at general availability.
Every cell below satisfies our eight-point Sealed Standard: both arms best-configured, workspace sealed, treatment verified on the transcript, raw runs published. Full methodology in /research.
Overall: 35–50% off agent spend on large codebases, and 2× fewer broken tests than baseline on SWE-bench-100. Nothing here is annualised or extrapolated. Below the size floor the saving thins and can invert; if your codebase is small and easy to grep, we will say so before you spend anything.
Fewer turns, fewer tokens, same result — measured on sealed paired tasks. Cost compounds across a session because context is re-billed every turn; Talon keeps it flat.
Founded February 2026. We optimise the loop your coding agent runs in; Talon is the first product: discovery, the right context in front of the model. Built for the way coding agents actually get used: long sessions, real codebases, and the cost and quality consequences that come with both.