Four labs, 48 hours, one architecture. What it costs, what’s oversold, and what to steal.
Between Wednesday morning and Friday afternoon this week, OpenAI, Cognition, DeepSeek and Mistral each shipped a piece of the same idea. None of them coordinated. By the end of this piece you’ll know what that idea is, what it actually costs to run, where the vendor numbers get shaky, and which parts you can borrow for your own work by Monday.
Verdict first: the frontier model is being promoted out of the work. The expensive, brilliant model now plans, delegates and reviews. Cheaper models, sandboxes and tools do the typing. If you’re still shopping for “the smartest model,” you’re shopping in the wrong aisle. The new question is who has the best system for handing the right job to the right intelligence at the right moment.
The org chart nobody drew on purpose
Anyone who has run a team already understands this. Back in my intel days, the senior analyst didn’t pull every image tile or georeference every scene. He wrote the tasking, the junior analysts ran the pulls, and he reviewed the product before it went out the door. Nobody paid him hourly rate to click through 400 tiles. That is also, almost exactly, how most of us have been using AI for two years. We hand the most expensive model on Earth a task and let it read every file, run every test and stare at every tool result itself. We’ve been paying the senior analyst to do data entry.
The receipts
OpenAI, September 10. The Agents API went into public beta. It’s the harness that runs Codex, rented out as a service: sessions that run for days, automatic context compaction so long jobs don’t drown in their own history, MCP tools, hosted sandboxes and parallel subagents.
Cognition, September 11. Devin Fusion arrived in Devin Desktop and the CLI. Fusion was first announced June 29; what’s new is that anyone can run it locally, and Cognition published benchmark runs done with Artificial Analysis and Vals AI. The design is the cleanest illustration of the whole thesis: a frontier “lead” owns the plan, the ambiguity and the final review. A cheaper “sidekick” explores the code, edits, runs tests and reports back. They pass briefs, results and feedback instead of sharing a conversation, and each keeps its own cached context. The lead never has to read the sidekick’s 40 intermediate tool results. Right? That alone is where a lot of the money goes.
DeepSeek, September 10. V4.1 Flash shipped as open weights under an MIT license. It’s a 552B mixture-of-experts model that only activates 16B parameters per token on output. Artificial Analysis ran it independently and, as of today, ranks it #6 of 113 comparable models on its Intelligence Index with a score of 40, versus 36 for DeepSeek’s own V4 Pro from August. Output speed on DeepSeek’s API measured 198 tokens per second. This is the worker in the stack: fast, cheap per token, downloadable.
Mistral, September 9. The most honest document of the week, and the least covered. Mistral’s applied team migrated the first 40,000 lines of a 300,000-line Fortran 77 reservoir simulator for a European energy operator. To document the codebase they spawned over a hundred agents through Vibe CLI, one per node of the call tree, each opening a pull request, with a reviewer agent on a cron loop checking PRs and scheduling fixes. Then they tried to let agents do the migration on full autonomy. The output was, in their words, Fortran retyped in C++ syntax. A planner, coder, tester and reviewer pipeline wrote better code, then stalled on bugs with no human around to unstick it. What worked was a human running a workflow of coder, tester and reviewer agents, module by module, with review gates before every merge.
Cognition also found that a more expensive lead can make the whole system cheaper. Their test swapping Opus 4.8 for Fable 5 as the lead: Fable costs twice as much per token, and Fable-led sessions came in 9% cheaper on average with a higher FrontierCode score, because Fable delegated earlier and wrote better briefs while Opus micromanaged the sidekick and redid its work. I’ve worked for that manager. So have you.
Claims vs. Receipts
Cognition says “up to 39% more efficient” (September 11). The receipt is the Artificial Analysis Coding Agent Index v1.5 run Cognition published the same day: Fusion with Fable 5.1 plus SWE-2 scored 61.7 at $7.90 per task, against Claude Code with Fable 5.1 at 62.2 for $12.36. That’s a 36% cost cut for half a point of score. Two caveats. AA and Vals were evaluation partners, and Cognition published the table, so treat it as vendor-published until a third party reproduces it. And the Astra pairing lost real ground on Terminal-Bench 4, dropping from 55.6 to 50.0 while saving 40%. Cheaper is not free.
DeepSeek’s Tianyi Cui said V4.1 Flash “comprehensively surpasses” V4 Pro (September 9, on X). Artificial Analysis partially backs that: Index 40 versus 36, and much faster. But the same page (checked September 11) shows V4.1 Flash burned 250 million output tokens to complete the Index, against a 140 million median for its size class, and lands at roughly $0.27 per Index task. The per-token price is low. The model talks a lot. Your bill is the product of both.
OpenAI’s launch quotes (September 10). Ciridae’s evaluation score going from 0.71 to 0.85 and SafetyKit’s 60% lower cost per case are real customer quotes, selected for the launch post. I found no independent Agents API benchmark as of Friday afternoon, so log it as promising.
“Mistral used 100+ agents to migrate Fortran.” That summary is half right. The hundred-plus agents documented the code. The migration itself succeeded only with humans at the gates, and 260,000 lines remain.
What it costs and who owns it
OpenAI Agents API: public beta, no separate fee during beta. You pay model tokens, tool usage and OpenAI container time at standard rates. Data residency is US-only and Zero Data Retention isn’t supported yet, which matters if you touch regulated data.
Devin Fusion: a mode inside Devin Desktop and CLI, so the cost is two models’ worth of tokens. The recommended pairing is Fable 5.1 as lead with SWE-2 (list price $0.75 per million tokens) as sidekick. Cognition notes SWE-2 costs 275% more per token than GPT-5.6 Luna and still came out slightly cheaper per task in their FrontierCode run.
DeepSeek V4.1 Flash: MIT license, weights on Hugging Face, so genuinely free to own. DeepSeek’s API listed at $0.30 per million input and $1.20 per million output on Artificial Analysis as of September 11. DeepSeek runs peak and off-peak pricing, so check the live page. Self-hosting still means a 552B-parameter footprint even with 16B active, so free means free after the GPUs.
Mistral: Vibe CLI on Pro or Team plans or an API key. The Fortran project was a services engagement, so the real cost there was engineer time.
So What
Do this: Split your next agent job into a planning role and a doing role before you pick models. Give the cheap model a written brief with success criteria and let the expensive one review. That’s the whole Fusion trick and it doesn’t require Devin.
Skip this: Full-autonomy swarms on anything you’d be embarrassed to explain. Mistral tried it with better engineers than most of us and got Fortran in a C++ costume.
Wait on this: Moving production workloads onto the Agents API until there’s an independent benchmark and until ZDR and non-US residency arrive.
Steal this line: Price per task is the only AI number that goes on an invoice.
Every figure above was pulled from the primary source or Artificial Analysis on September 11, 2026. Prices and benchmarks here move weekly, so re-verify before you quote them.
Sources
OpenAI: Introducing the Agents API (September 10, 2026)
Cognition: Introducing Fusion in Devin Desktop & CLI (September 11, 2026)
Cognition: Devin Fusion original announcement (June 29, 2026)
Artificial Analysis: DeepSeek V4.1 Flash model page (checked September 11, 2026)
Mistral: Modernizing complex legacy code with AI agents (September 9, 2026)
The Decoder: New DeepSeek model V4.1 Flash cuts memory needs for AI agents
RuntimeWire: DeepSeek plans V4.1 Flash for September 10 and will route Pro traffic to it
AlphaSignal: Cognition’s Fusion cuts Devin coding agent costs
arXiv: Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents (referenced in production notes)