Gemini 4 Argon and Claude Opus 5.5 make agent work cheaper

Look at the AI coding models 2026 has produced so far and one trend stands out: long, multi-step engineering work is getting cheaper and faster. Anthropic released Claude Opus 5.5 on September 22 ^[3], and Google announced Gemini 4 Argon eight days later, on September 30 ^[1].

Both companies pitch their model at the same job: agents that work for hours on migrations, refactors and research, not quick autocomplete ^{[2, 3]}. For small product teams and studios like ours, that shifts which projects are worth taking on. This piece covers what each model offers, what the numbers do and do not tell you, and how we suggest a design and development team responds.

Key takeaways

• Gemini 4 Argon raises the output limit to 1 million tokens, up from 64K, at an introductory $2 per million input tokens and $10 per million output tokens ^[1].
• Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, and Anthropic says it runs 40% cheaper than Opus 5 on typical workloads ^[3].
• Each company reports different benchmarks, so the scores do not compare directly ^{[1, 3]}.
• The bottleneck for teams moves from writing code to specifying and reviewing it.

Gemini 4 Argon: built for long outputs

Google describes Gemini 4 Argon as its frontier model for complex workflows across software engineering, enterprise knowledge work and cybersecurity defense ^[2]. The headline change is output length. Argon can produce up to 1 million tokens in one response, up from the previous 64K ^[1]. Google’s developer documentation adds a detail worth knowing: the default synchronous output is 65,536 tokens, and the full 1,048,576 tokens requires the long decode mode ^[4]. The context window is also 1,048,576 tokens ^[4].

Pricing launches at an introductory $2 per million input tokens and $10 per million output tokens, with cached input priced at 95% off ^[1]. After the introductory period, Google says the price rises to $4 and $20 ^[1]. Google reports a state-of-the-art 77.9% on DeepSWE v1.1, a benchmark for long-horizon software engineering tasks ^[1].

Developers have been able to build with Argon through the paid Gemini API tier in Google AI Studio and Android Studio since October 13 ^[2]. The model accepts text, image, video, audio and PDF input, outputs text, and supports function calling, code execution, structured outputs and a preview of computer use ^[4].

“Gemini 4 Argon is designed for those long-horizon workflows that require deep reasoning and structural precision, including code migrations and frontier research.”

Claude Opus 5.5: cheaper and faster than Opus 5

Anthropic released Claude Opus 5.5 on September 22 at $4 per million input tokens and $20 per million output tokens ^[3]. The company says it costs 40% less than Opus 5 on typical workloads and generates output 30% faster ^[3]. It is available on Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure ^[3].

On Anthropic’s own numbers, Opus 5.5 scores 66.4% on Terminal-Bench 4.0, up from 52.3% for Opus 5, and 57.8% on CursorBench 4.0, up from 46.6% ^[3]. Anthropic’s documentation lists a 1 million token context window, a 128K token output limit on the standard Messages API, and a June 2026 knowledge cutoff ^[5]. On the Message Batches API, output can reach 300K tokens with a beta header, and batch requests are 50% off ^[5].

“One tester completed a 680,000-line code migration in less than a day—work that would have taken an engineering team weeks.”

Independent coverage points the same way. DeepLearning.AI’s The Batch reports that Opus 5.5 scored 58 on the Artificial Analysis Intelligence Index v4.3, seven points above Opus 5 ^[6].

Comparing AI coding models 2026: read benchmarks with care

It is tempting to line the two models up in one table and pick a winner. The data does not support that. Google leads with DeepSWE v1.1 ^[1]. Anthropic leads with Terminal-Bench 4.0 and CursorBench 4.0 ^[3]. These test different things under different setups, so a higher number on one says little about the other.

Price comparisons need the same care. Argon’s $2 and $10 is an introductory rate, and Google has already published the $4 and $20 it moves to afterwards ^[1]. That later price matches Opus 5.5’s list price ^[3]. Caching and batch discounts change the real bill a lot on both platforms ^{[1, 5]}, so the cheapest model on paper may not be the cheapest for your workload.

What does compare is the direction. Output limits are longer, list prices are flat or falling, and both vendors now market their flagship models on multi-hour engineering tasks rather than chat ^{[1, 3]}. Work that used to be too expensive for a small studio, such as framework migrations, test coverage for legacy code or a design-token refactor across hundreds of files, is now realistic to scope.

What changes for designers and developers

When an agent writes most of the diff, the scarce skill is no longer typing code. It is writing a clear brief, knowing what good looks like, and catching the mistake that passes the tests but breaks the product. That puts more weight on work designers and senior developers already do.

• Specs get more valuable. An agent follows written rules well and guesses badly. Design tokens, component rules and acceptance criteria are now direct inputs to the build.
• Tests become the contract. A migration an agent finishes in a day is only safe if a test suite can prove nothing broke.
• Review becomes the job. Someone who knows the product has to read the output, check edge cases and own the merge.
• Design QA moves earlier. If screens get built in hours, visual and accessibility checks need to run in the same loop, not at the end.

How your team should act now

Start with one contained project, not a platform switch. A good first candidate is a migration or refactor with clear tests: upgrading a framework version, moving styles to tokens, or adding types to an untyped codebase. Measure cost, time and the number of review comments, then decide whether to scale.

Keep your setup model-agnostic. Both models are available through several clouds ^{[2, 3]}, and prices change: Argon’s already has a published increase ^[1]. Wrap model calls behind one interface and keep prompts and specs in your repository, so switching is a config change rather than a rewrite.

Finally, budget for review. The time an agent saves on writing should go into reading its work, not into taking on twice the scope with the same reviewers.

FAQ

Is Gemini 4 Argon cheaper than Claude Opus 5.5? For now, yes, on list price. Argon’s introductory rate is $2 input and $10 output per million tokens, against $4 and $20 for Opus 5.5 ^{[1, 3]}. Google says Argon moves to $4 and $20 after the introductory period ^[1].

Which model is better for coding? The vendors report different benchmarks, so there is no direct answer from their data ^{[1, 3]}. Test both on a real task from your own codebase.

Can these models really handle a full code migration? Anthropic reports one tester finished a 680,000-line migration in under a day ^[3], and Google names code migrations as a target use for Argon ^[2]. Treat these as vendor claims and keep strong tests and human review in place.