What happened
GIGAZINE reports that Cognition, the company behind Devin, released Fusion, a harness designed to improve how advanced AI models plan, execute, review, and recover from stalled tasks. The system pairs a high-end model for planning and review with a less expensive model for execution, while running two agents in parallel with separate contexts and tools.
GIGAZINE reports that Cognition developed and released Fusion as a harness for OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable. The article describes a harness as the layer that supplies models with operating instructions, tool-use rules, and conditional branching, potentially affecting whether an agent completes a task or becomes stuck in a loop.
According to GIGAZINE, Fusion combines a state-of-the-art model for planning and review with a cost-effective model for execution. Cognition recommended GPT-6 Astra and Claude Fable as the advanced models, and identified Cognition’s SWE-2 coding model as the cheaper execution option. The article says the combination performed best with Claude Fable 5.1.
GIGAZINE reports a benchmark score of 62.2 for Claude Fable 5.1 with Claude Code, compared with 61.7 for Claude Fable 5.1 combined with the lower-cost model and Fusion. The latter combination was reported as 36% cheaper. For GPT-6 Astra, the article says Fusion achieved performance equivalent to Codex while reducing costs by 39%.
The article attributes evaluations to AI analytics companies Artificial Analysis and Vals AI. It says Fusion runs two agents in parallel, each with independent context and tools, and exchanges only summaries, results, and feedback rather than the full conversation. GIGAZINE says this approach makes greater use of prompt caching. The source does not provide the full benchmark methodology, absolute prices, or availability terms.
Source details: gigazine.net ↗
Why it matters
If the reported results hold, Fusion could reduce the cost of using frontier AI agents without a comparable loss in coding performance. That matters for developers and organizations running tool-using systems at scale, where model inference and repeated agent attempts can become major expenses. The results are reported by GIGAZINE and evaluated by Artificial Analysis and Vals AI, but the source does not provide enough methodology to independently assess the comparisons.
The reported results suggest that agent architecture can materially affect the economics of frontier models. A system that delegates execution to a cheaper model while retaining a stronger model for planning and review could lower the cost of software agents without requiring users to abandon higher-capability models.
Cost reductions are particularly relevant for coding agents, which may make many tool calls, repeat failed attempts, or maintain long contexts. If independently reproducible, the reported 36% and 39% reductions could influence how companies design production agent workflows and choose between proprietary harnesses.
The evidence remains limited in the supplied report. GIGAZINE provides selected scores and percentage savings but not the task set, run count, pricing assumptions, latency, failure rates, or comparison conditions. The claimed equivalence to Codex and the stated efficiency gains are therefore not independently confirmed here.
What to watch next
The key questions are who can access Fusion, which models and tools it supports, whether it is generally available, and how its costs are calculated. Further reporting should clarify the benchmark tasks, baselines, sample sizes, prompt-cache assumptions, and whether the results transfer beyond coding. The source also uses both “SWE-2” and “SWR-2” for the lower-cost model, an inconsistency that should be resolved.
Access conditions are not documented in the source. It is unknown whether Fusion is publicly downloadable, available through Cognition’s products, restricted to selected users, or offered through another commercial arrangement. Pricing and any required model subscriptions are also unknown.
Follow-up reporting should establish whether Fusion supports models beyond GPT-6 Astra and Claude Fable, whether users can supply their own tools and prompts, and how much of the savings depends on prompt caching or specific provider pricing.
The benchmark naming contains a source-level inconsistency: the article identifies the cheaper coding model as SWE-2 but later refers to “SWR-2.” That designation should be verified before treating the comparison as a precise product claim.
Independent testing should examine coding success rates, latency, reliability, context handling, and performance on tasks other than the reported benchmark. Parallel agents may also increase orchestration complexity or total tool activity even when model costs fall.