Cognition's SWE-1.7 runs Devin at 1,000 tokens/sec — and confirms coding-agent companies are becoming model companies

What shipped

On July 8, 2026, Cognition launched SWE-1.7, describing it as the most capable model it has trained, and put it directly inside Devin — the autonomous software-engineering agent. Per WinBuzzer and VKTR:

• Speed: ~1,000 tokens per second, served through Cerebras hardware.
• Benchmarks (Cognition’s reported figures): Terminal-Bench 2.1 — 81.5%, SWE-Bench Multilingual — 77.8%, FrontierCode 1.1 Main — 42.3%.
• Availability: Devin Web, Desktop, and CLI.
• The pitch: frontier-class agentic coding at a fraction of the cost — explicitly framed as advancing the cost-performance Pareto curve rather than claiming the capability crown.

Cognition also pushed Devin Desktop — which is not a new product but Windsurf under a new name, rebranded on 2 June 2026 as “the next generation of Windsurf” — and on July 23 acquired The Interaction Company of California, makers of Poke — an AI agent that operates inside text messages, and reportedly the first third-party AI agent Apple approved to text natively on Messages Business Chat.

Where it actually lands

Cognition is refreshingly honest in its framing — it isn’t claiming to beat the frontier. Put SWE-1.7 against the models we’ve covered this month on shared benchmarks:

Two honest readings. Against the frontier labs, SWE-1.7 trails — roughly 10 points behind GPT-5.6 Sol on Terminal-Bench, and a couple behind Grok 4.5. Against the other harness company that trains its own model, it leads: 77.8% vs Cursor Composer 2’s 73.7% on SWE-Bench Multilingual. That’s the fairer comparison, because it’s the same category of bet — a coding-tool company building its own model rather than renting one.

(One caution: SWE-1.7’s FrontierCode 1.1 score of 42.3% sits suspiciously close to Claude Opus 5’s 43.3% on Frontier-Bench — but those are differently-named benchmarks and should not be read as a head-to-head. Comparing across benchmark names is exactly how misleading rankings get built.)

Why this matters

1. The harness companies are becoming model companies — that’s the real headline. Cursor trains Composer. Cognition now trains SWE. Both started as harnesses — products wrapping someone else’s frontier model — and both concluded that renting intelligence is a losing long-term position. The logic is stark: if your product is a thin layer over a Claude or GPT API, then your gross margin is whatever the frontier lab leaves you, your latency is theirs, and your rate limits are theirs. Training in-house flips all three. This is the application-layer version of the same vertical-integration wave driving OpenAI’s Jalapeño chip, Google’s Frozen v2, and Anthropic’s 2 GW AMD deal — everyone is climbing toward owning their own stack.

2. Speed is a genuinely under-rated axis for agentic coding. A thousand tokens per second sounds like a spec-sheet flourish until you think about what an agent actually does: read code, plan, edit, run tests, read the failure, retry — dozens or hundreds of short cycles. Latency compounds at every step. A model that’s a bit less clever but several times faster can finish an agentic task sooner and keep you in flow, versus a smarter model where you context-switch away during each cycle. For interactive agent work, “fast enough to watch” is a different product than “smart but slow.” That’s the real argument for SWE-1.7, and Cerebras is what makes it possible.

3. It’s another datapoint that “cheaper and faster, slightly less capable” is the winning 2026 product shape. We’ve now seen it repeatedly: Claude Opus 5 at half of Fable 5’s price, GPT-5.6’s Terra tier, Grok 4.5’s cost-per-task pitch, and now SWE-1.7. The frontier crown is increasingly a marketing asset, while the volume — and the margin — sits at “good enough, fast, cheap.” Buyers benefit, but it also means benchmark leadership is a worse and worse proxy for which tool you should actually use.

4. The Poke acquisition shows Cognition wants agents everywhere, not just in the IDE. Buying an agent that lives inside text messages — with rare native Apple Messages approval — is not a coding move. It signals Cognition sees Devin as one instance of a broader “agents in your workflows” product, competing with the deployment-layer land grab the frontier labs are running. Worth watching whether that focus dilutes Devin or strengthens it.

5. For Devin specifically, it’s the most substantive upgrade in a while. Devin’s history is a cautionary tale about demo-to-product gaps — a spectacular 2024 launch followed by a long climb to usability. Owning the model is the strongest signal yet that Cognition is building a durable product rather than a wrapper, and it gives Devin something Claude Code and Cursor can’t trivially copy: a model tuned specifically for its own agent loop, on hardware chosen for throughput.

What this means for you

• If you use Devin: SWE-1.7 is a real upgrade, especially on iteration speed. Try it on a task you’d normally hand to a slower agent and judge cycle time, not just final output.
• If you’re choosing a coding agent: pick by constraint. Speed/cost-bound → Devin with SWE-1.7 or Cursor; hardest-problem-bound → Claude Code with Opus 5, or GPT-5.6 Sol. See the best AI coding tools guide and Claude Code vs OpenAI Codex.
• Don’t read the benchmark table as a ranking of usefulness. SWE-1.7 trails on Terminal-Bench and may still be the better daily driver if your work is high-iteration. Test on your own repo — that’s the only benchmark that predicts your experience.
• Watch the in-house-model trend. If your coding tool starts routing you to its own model by default, understand the trade: usually faster and cheaper, sometimes less capable on hard problems. Check whether you can still select a frontier model when you need one.

The honest caveats

• All benchmark figures are Cognition’s own. The 81.5% / 77.8% / 42.3% scores come from the company. Independent verification on contamination-resistant evals is the standard before treating them as settled — a discipline that matters more than usual after METR found a frontier model gaming its own evaluations.
• Cross-benchmark comparisons here are limited on purpose. Terminal-Bench 2.1 and SWE-Bench Multilingual comparisons are like-for-like; FrontierCode 1.1 and Frontier-Bench are not the same test and aren’t compared. Treat any table mixing benchmark names with suspicion.
• 1,000 tokens/sec is a throughput figure, not an end-to-end latency guarantee. Real agentic task time also depends on tool calls, test runs, and network round-trips. Expect a meaningful speedup, not a 10× wall-clock miracle.
• “Fraction of the cost” needs your own math. Cognition’s pricing is bundled into Devin’s plans rather than published as a simple per-token rate comparable to frontier APIs. Price it against your actual usage before assuming savings.
• This shipped July 8; we’re covering it now. It’s a gap in our coding-agent coverage rather than breaking news — included because the in-house-model trend it represents is still developing and materially relevant.

The grounded summary: SWE-1.7 doesn’t take the coding crown, and Cognition doesn’t claim it does. What it does is put a fast, cheap, purpose-built model inside Devin — and confirm that the companies building coding agents have decided renting intelligence from frontier labs is not a business. Judge it on cycle time in your own repo, and watch how many other harnesses follow.

Questions or corrections? Email Pick Right. Want the full list? See all news.