Devin Fusion [What It Changes]

TL;DR

• Devin Fusion is Cognition's multi-agent harness that keeps a frontier orchestrator in charge and delegates mechanical work to a cheaper "sidekick" agent with its own cached context.

• Cognition listed Fusion among June 2026 ships. The product post reports FrontierCode 1.1 Extended scores with Fusion at 63.1 and $1.35 per task versus Fable 5 at 64.9 and $10.53 (data updated 8/7/2026).

• Fusion is a Devin architecture, not an open protocol. Do not treat 60.7 vs 54.6 (AIToolsReview's Fable-vs-Opus Fusion study) as a public leaderboard.

• A two-truck HVAC shop, a 10-person agency, or a solo clinic should care when the contractor bills frontier rates to run the test suite after a one-line fix.

One writer lock

Fusion as an orchestrator over coding agents fails when five agents edit one file. Two agents first. Single writer lock. Conflicts into review, not into main.

Orchestrator with a lock

Devin Fusion as one orchestrator over a team of coding agents needs a single writer lock. Five agents on one file is a merge conflict factory. Two agents first. Conflicts into review. Source pack for Fusion. No invented Cognition price. Do not copy a ten-agent demo into a ten-person shop.

Partner memo for Devin Fusion [What It Changes]

The empty object is the only decision. Write it in one sentence on the whiteboard. If you cannot, you are still in a demo.

Quotes are dated PDFs. "Around" is still a figure we will not print unless the brief's price policy allows it with an ISO date on the same line.

Week one: kill one shadow path — a personal phone, a second login, or a spreadsheet that is pretending to be the record. NFIB's 2024 figure of 44% of small businesses citing time-management as a top challenge is why you do not migrate two systems in the same sprint.

Week two: one named owner for failures. If the owner is "whoever built it," you do not have an owner.

Week three: count the copy-paste jobs that remain. That count is the workflow, not a reason to smash two products into one license.

SBA's 2025 profile of 33M+ small businesses includes shops that bought both logos and finished neither. Sign one quote. Schedule the rest 60 days later.

F518 lives or dies on whether that sentence on the whiteboard matches the screen staff will actually live in. If the screens disagree, you picked the demo, not the leak.

Close-out checklist for Devin Fusion [What It Changes]

1. Dated quote in the folder, or a written "quote only" if no public figure exists.

2. Named owner for week-one failures, not "the founder when they see it."

3. One shadow path killed: personal phone, second login, or spreadsheet-as-record.

4. Internal links in this page still resolve on the live site; homepage is https://ustechautomations.com/.

5. No second product in the same sprint. NFIB 44% is the constraint.

If any line is unchecked, you are not live. You have a login. F518 should not ship a second logo until those five lines are true. SBA's 33M+ small businesses include a lot of logins. Be the shop that finished one object.

Goldman Sachs' 62% self-reported workflow ROI inside 12 months starts when the old path is dead, not when the demo ended. Kill the old path. Then stop.

Desk rule for Devin Fusion [What It Changes]

Source pack first. No invented vendor price. One shadow path killed this week. Humans keep merge rights. If the run is still on a personal login, it is not a desk tool. Pin the output to the job in the record. If you cannot name the record, stop.

Two agents first. One writer lock. Fusion is not a reason to skip review.

Date the decision for Devin Fusion [What It Changes]. If the PDF has no date, you do not have a comparison. Kill one shadow path this week. Do not add a second logo until the first object is true. NFIB 44% is why the second sprint waits.

According to AICPA, 62% of firms reported cloud-workflow adoption.

According to Journal of Accountancy, the mid-market close still runs 8-10 business days.

According to Thomson Reuters, tax-prep utilization hits 85-95% in March and April.

According to NFIB, 44% of small businesses cite time-management as a top challenge.

According to NFIB, 44% of small businesses cite time-management. According to SBA Office of Advocacy, 33M+ small businesses sit in the 2025 profile. According to Goldman Sachs, 62% of SMBs reported workflow-tool ROI inside 12 months.

Key Takeaways

• Sidekick + dynamic mid-session routing are the two techniques Cognition names.

• An internal Fusion preview found 88% of merged PRs driven entirely by the automated Fusion router.

• Fable 5 access was suspended 12 June 2026 per a U.S. government directive; Fable numbers are from before that cutoff.

• AIToolsReview independently reports the Fusion study (Fable 5 vs Opus 4.8 as orchestrator) as Cognition-stated.

• Human merge still applies. Fusion does not make Devin an unsupervised production deploy.

What Devin Fusion is, in one sentence

Devin Fusion is a Devin harness that runs a frontier lead agent and a cheaper sidekick agent in parallel, each with its own tools and cached context, and routes work between them mid-session. That is the entity. It is not SWE-1.7 (the in-house model), not Security Swarm, and not ACP.

A two-truck HVAC shop does not pick an orchestrator. It pays a person who leaves an agent on a booking-form bug. If the expensive model also runs the Playwright suite, the bill is a Lamborghini to the grocery store — Cognition's own analogy. A 10-person agency hits that on a CSS tweak. A solo clinic hits it on a portal copy change. Fusion is the routing layer that is supposed to keep judgment on the expensive model and chores on the cheap one.

Shops comparing Clio alternatives already know a suite SKU is not a new system of record. Fusion lives inside Devin. Smokeball vs Clio and MyCase vs Clio are still the matter-system choice; Fusion is how the vendor's coding agent spends tokens.

Why a small operator should care before the deep dive

According to the SBA Office of Advocacy, the United States contains 36.2 million U.S. small businesses. Almost none will tune a dual-agent harness. They still pay for overnight runs.

According to Cognition's Fusion post, Fusion maintains frontier and Fable 5-level performance at up to 60 percent lower cost on FrontierCode (dagger: originally 35 percent at publication; updated 8/7/2026). That is Cognition's cost claim.

Cognition's funding post states a $26 billion valuation, over $1 billion raised, and $492 million run-rate. The anniversary post lists Devin Fusion in June 2026 next to FrontierCode.

If the office already automates assistant chores as in five ways to automate assistant tasks, Fusion is that split: judgment vs grunt work. Teams already routing those chores through US Tech Automations workflows can treat Fusion as a routing swap on the same ticket, not a rebuild of intake. Form capture still sits in form-to-CRM automation.

What Cognition shipped in June 2026

The Fusion post is the product write-up. Preview signup: app.devin.ai/signup. FrontierCode is the quality benchmark Fusion is scored on.

Cognition's Extended chart (updated 8/7/2026):

Source: Cognition Devin Fusion post. Vendor chart, not a public leaderboard.

Without Fable 5, Fusion is claimed at up to 60 percent cost improvement while maintaining frontier-level performance. With Fable 5, Fusion is 41 percent cheaper than a pure Fable 5 harness, matching Fable 5's performance in a traditional harness. Fable 5 access was suspended 12 June 2026 (Anthropic notice); those Fable numbers are pre-suspension.

AIToolsReview restates a narrower Fusion study on FrontierCode 1.1:

Source: AIToolsReview July 2026, summarizing Cognition. Do not treat 60.7 vs 54.6 as an independent ranking.

AIToolsReview adds Cognition's behavioral explanation: Fable 5 averaged 11.5 lead-agent turns vs Opus 26.5, and never made a direct code edit in 81 percent of runs vs 24 percent for Opus-led runs.

Worked examples in the Fusion post (sidekick on vs off): search.js modernization −62% cost ($3.55 → $1.37); OpenTracing removal −32% ($3.80 → $2.57); JSON-Schema oneOf −38% ($5.08 → $3.13); hard React/Redux feature −28% cost but score collapse 75 → 27 when judgment was delegated; LangChain4j/Quarkus −25% ($5.25 → $3.93) with a score gain.

Prior art Cognition cites: its own "Smart Friend" tool and Anthropic's "Advisor" tool — both pay a cache miss when the other model is queried. Sidekick keeps two persistent cached contexts. Cached inputs are noted as having a 5-minute expiry.

A NYT piece on token-minimizing is linked from the Fusion post as the cost backdrop. Open-source model mixing is pointed at a Kimi / GLM Devin Desktop post.

How sidekick and routing actually work

Two agents run. The lead should take few actions, delegate, monitor, and keep the plan, the ambiguous calls, and the final review. The sidekick does mechanical work in its own context. Classifiers during the run can promote work back to the lead or switch models. Cognition switches models at context compaction so the cache miss was happening anyway.

The constraint that broke is "one model for the whole session." Prompts do not encode difficulty well, and a cheap first prompt can get a hard follow-up.

The honest limit: Fusion is Devin's harness. It is not a spec another IDE can implement. Bad delegation on a judgment-heavy task (the React example) drops the score. Humans still merge.

NIST AI RMF remains the voluntary U.S. vocabulary if a shop needs to write down that routing is in use. NIST does not score Fusion.

Internal preview: 88 percent of merged PRs driven entirely by the automated Fusion router — Cognition dogfood, not a customer sample.

USTA analysis: $10.53 vs $1.35 on the Extended chart

USTA analysis. Working only from figures already cited above: Fable 5 at $10.53 and score 64.9; Devin Fusion at $1.35 and 63.1 (Extended chart, 8/7/2026).

• Dollar delta: $10.53 − $1.35 = $9.18 per task.

• Cost ratio: $1.35 / $10.53 ≈ 12.8% of Fable 5's cost, i.e. about 87.2% cheaper on this row — larger than the "up to 60%" headline, which Cognition applies to a broader frontier comparison, not solely this Fable 5 cell.

• Score delta: 64.9 − 63.1 = 1.8 points.

• AIToolsReview's Fable+sidekick $1.86 vs standalone $4.03 is a different table (FrontierCode 1.1 study): $4.03 − $1.86 = $2.17, or 53.8% cheaper, with scores 60.8 vs 60.7.

Sources for inputs: Fusion post; AIToolsReview. Results are USTA arithmetic on vendor numbers, not a public leaderboard.

Do not mix the Extended chart (63.1 / $1.35) with the 60.7 / $1.86 study as one score.

What this does not establish

Fusion does not establish an open protocol. It does not establish that Fable 5 is available. It does not establish that 88% router-driven PRs will hold at a customer.

It does not establish a public Fusion SKU price. Devin plans in the July roundup (Pro $20, Max $200, Teams $80+$40) are the platform, not Fusion à la carte. Check the vendor site.

The 60.7 vs 54.6 pair is Cognition's study on Cognition's harness. AIToolsReview says so.

A buyer's evaluation sequence

None of the sourced posts say Fusion merges unsupervised. Use four checkpoints.

A startup that wants a generic agentic path can look at agentic workflows. US Tech Automations is the logging and approval layer if a Fusion run still needs a human hold before anyone deploys.

Pricing on this site is that layer. The wider picture is the state of small-business automation.

Signal vs Speculation

Demonstrated signal: as of June 2026 Cognition listed Devin Fusion among ships; the Fusion post gives the Extended cost/score chart (updated 8/7/2026), 60% / 41% cost claims, 88% internal router share, sidekick examples with percent cost cuts, and Fable 5 suspension on 12 June 2026; AIToolsReview independently restates the Fable-vs-Opus Fusion study (60.7 vs 54.6, $1.86 vs $2.04) as Cognition-stated; anniversary and Series D posts supply company scale ($26B, $492M); Fusion is a Devin architecture, not an open protocol.

Our read: over the next 12–36 months, small shops will not "install Fusion." They will notice whether the contractor's agent uses a cheap model for tests and an expensive one for the design call. If Cognition's 60% cost line holds on chores and fails on judgment tasks, the practical contract language is: route tests and mechanical refactors to the sidekick path; keep auth, payments, and PHI-adjacent changes on a named frontier model with a human merge. Do not wait for a portable fusion protocol; these sources do not describe one.

Fusion is one orchestrator over coding agents

Devin Fusion letting one orchestrator run a team of coding agents is a manager process. The failure is five agents editing one file. Require a single writer lock.

Signal: Fusion launched. Speculation: every coding desk copies it. Start with two agents, not ten. US Tech Automations can monitor the orchestrator output and flag conflicts into the review queue.

FAQ

What is Devin Fusion?

A Devin harness with a frontier lead agent and a cheaper sidekick agent, plus mid-session routing.

When did it ship?

Cognition's anniversary list puts Fusion in June 2026. The product post is the detail surface.

Is 60.7 vs 54.6 a public ranking?

No. It is Cognition's Fusion study as restated by AIToolsReview.

Is Fable 5 available?

Cognition says access was suspended 12 June 2026 under a U.S. government directive and had not been restored at the time of the Fusion post.

Does Fusion replace SWE-1.7?

No. SWE-1.7 is a model SKU. Fusion is a routing harness that can mix models.

What should a small shop put in a contractor agreement?

Require a cheap path for tests, a named frontier model for judgment, and a human merge. Ask to see the bill split, not only the PR.

What to do next

Devin Fusion is a Devin-only orchestrator-plus-sidekick harness with vendor cost charts. The honest limit is the self-reported studies and the Fable 5 cutoff.

If the next step is to put that routing behind an approval step instead of letting the cheap path touch payments code, start from agentic workflows for Fusion-style model routing on the wiring layer, then set the hold on pricing.