Devin Gets 30–40% Cheaper as Fusion Tops FrontierCode 1.1 With a 68.8 Score

Devin has become noticeably cheaper. According to the developers, prices fell by 30–40% in Fusion and Normal modes, by 15–20% in Ultra, and by up to 70% in Devin Review. The team claims intelligence has been preserved or improved in every mode. Devin Fusion took first place on FrontierCode 1.1 Extended with a score of 68.8 at an average cost of $0.60 per task.

A cheaper agent and a stronger agent usually require a trade-off. Here, the authors claim to have achieved both at once. The number in the table looks convincing, but the measurements were published by the vendor itself, and the comparison includes its own SWE-2 model. Test it on your own tasks.

Models: matching the model to the task

Models account for two-thirds of the savings, and the engineering around them delivers the rest. Devin is not tied to a single provider and picks a model for each stage of work. According to the team:

• Opus 5.5 and GPT-6 Sol offer stronger intelligence at a good price-to-quality ratio.
• GPT-6 Astra handles computer use better than the others.
• GPT-6 Luna noticeably cuts the cost of auxiliary tasks.
• SWE-2 is the team's own model.

Fusion is built on a pair: a capable lead model and a cheaper helper. Devin Review selects models separately for diff analysis, bug hunting, and change categorization. Normal and Ultra choose them for planning, testing, debugging, and migrations.

Benchmark: score versus price

FrontierCode 1.1 Extended, average cost per task and score:

• Devin Fusion: 68.8 points, $0.60.
• Opus 5.5 (high): 65.2, $0.90.
• Fable 5.1 (medium): 63.6, $2.68.
• GPT-6 Astra (high): 63.1, $2.62.
• SWE-2 (high): 60.1, $0.64.
• GPT-6 Sol (high): 59.6, $0.87.

Fusion beats Opus 5.5 by 3.6 points and costs a third less. Against Fable 5.1 and GPT-6 Astra, it is more than four times cheaper. Also worth noting: SWE-2 on its own costs barely more than Fusion but trails it by 8.7 points. The pairing of models is what delivers the quality gain.

Harness: fewer turns, more cache

The second source of savings is the harness, the layer that controls the agent loop. The new models are better at planning several actions at once. Previously, formatting, linting, and tests ran as three separate turns. Now the model can request all three in a single call and reason about the results together. In the team's illustrative example, that means 2 turns instead of 4 and 49% fewer tokens sent. Independent calls now run in parallel, and planned sequences run back to back without returning to the model.

The second part involves prompt caching. LLMs are stateless, so every agent request carries the full history: instructions, code, and past actions. If the beginning of a request matches the previous one, the provider reuses the computation for the shared prefix. In a demo session of five turns, the cache cuts the tokens processed from scratch to 27K instead of 93K, which is 71% fewer. These are also schematic, rounded figures, not a measurement on real tasks.

The approach has limitations. The shared prefix must stay unchanged, the cache is short-lived, and it is tied to a specific model and provider. That means the harness has to be designed carefully. On the other hand, it benefits from cheaper cache reads: for Opus 5.5 they cost 60% less than for Opus 5, and GPT-6 Sol and Luna cost half as much as their predecessors in the 5.6 line. In long sessions with a large history, these discounts add up.

What teams should do

If you already pay for Devin, it makes sense to recalculate your budget: the same tasks now cost less, especially in Review. If automated pull request review was too expensive, a discount of up to 70% changes the math. Run a couple of typical tasks on your own repository in Normal and Fusion, compare the total cost and the share of accepted changes, and only then switch the default mode.

For those building their own agents, the lesson is more practical. Keep a stable request prefix so the cache works. Let the model batch commands into a single call. Assign cheap models to auxiliary steps. The update is already available at devin.ai.