Gemini 4 coding results tell a mixed story. Google says its new Argon model leads several software tests, yet its own table shows losses on other engineering tasks. Bloomberg also reports doubts from some employees who used it. Google disputes the suggestion that coding is a weakness. For buyers, the missing evidence is an independent test on their own work.

Key takeaways

  • Google announced Gemini 4 Argon on 30 September 2026, with initial access limited to selected cyber defenders.
  • Google reports 77.9% on DeepSWE v1.1, above the peers in its table, but 55.0% on FrontierSWE v2, below three listed rivals.
  • Bloomberg cited people with access who questioned some practical coding performance. Google rejected a broad underperformance label.
  • No public result here proves that one model is best for every development team. A representative, supervised trial remains essential.

Google’s 30 September launch announcement and DeepMind benchmark table establish what the company has actually published. Bloomberg’s original report describes a different signal: concerns from people close to the rollout about certain everyday coding tasks. These are different kinds of evidence. The first consists of vendor-reported test results. The second is attributed reporting about experience inside one organisation. Neither is a controlled comparison of the model on an Indian company’s codebase.

What the Gemini 4 coding numbers actually say

Four coding rows in Google’s table are especially useful because they do not all point in the same direction. On DeepSWE v1.1, Google lists Argon at 77.9%, GPT-6 Astra at 74.1%, Claude Fable 5.1 at 67.4%, and Claude Opus 5.5 at 74.2%. On Vibe Code Bench, Google reports 91.9% for Argon, just ahead of the peers shown. Those figures support a claim of strong performance under Google’s stated conditions. They do not establish universal superiority.

The same table puts Argon at 55.0% on FrontierSWE v2. That is below Google’s listed 65.5% for Astra, 62.3% for Opus and 56.3% for Fable. On Terminal-bench 4.0, Google lists Argon at 57.4%, behind 66.4% for Opus, 58.2% for Astra and 57.9% for Fable. The important point is not to declare a winner from one row. It is that benchmark choice changes the ranking even within the vendor’s own release material.

Google-reported coding scores for Gemini 4 ArgonDeepSWE v1.1 77.9 percent; FrontierSWE v2 55 percent; Vibe Code Bench 91.9 percent; Terminal-bench 4.0 57.4 percent. These are different tests and should not be averaged.Argon across four different testsVendor-reported percentages; tests are not interchangeableDeepSWE v1.1FrontierSWE v2Vibe Code BenchTerminal-bench 4.077.9%55.0%91.9%57.4%Source: Google DeepMind, 30 September 2026

Google’s scores are attributed to Google because the company selected and published this comparison. They are not described here as independently reproduced. The accompanying methodology matters: prompts, tools, time limits and scoring rules can change outcomes. A reader should also avoid averaging the four percentages, because they measure different tasks. A percentage on a software repair benchmark does not mean the same thing as one on a test of generating an interface from a brief.

What Bloomberg reported, and what Google denied

Bloomberg’s report, syndicated on Yahoo Finance, said some people with direct access found that Gemini 4 did less well when employees put it to work than its benchmark position suggested. The report singled out some coding work, including front-end design. Bloomberg also reported a contrasting view from a Google employee who said there was broad internal agreement that the model was at the frontier. That makes the account a report of disagreement, not a finding that the system fails at coding generally.

Google told Bloomberg that it would be inaccurate to say Gemini 4 underperforms in coding. The company’s announcement points to specialised code tasks and internal examples of longer engineering work. Axios independently covered the limited launch and noted the Bloomberg criticism. Ars Technica examined the stated software scores and access restriction. VentureBeat reviewed Google’s release material and called out its weaker rows. These outlets offer separate launch reporting and analysis; their references to Bloomberg do not turn one employee-sourcing exercise into several independent employee interviews.

The distinction matters for a responsible headline. “Some employees question particular coding tasks” is supported by the Bloomberg account. “Gemini 4 cannot code” is not. It would also be wrong to turn Google’s strongest test score into proof that it outperforms every rival on all software work. We have kept those claims separate so readers can see which observation came from the company, which came from reporters, and which remains untested outside the early-access group.

Why a coding benchmark is a limited buying signal

Software work includes bug fixes, system design, tests, documentation, user interfaces, deployment and maintenance. An evaluation usually measures a narrower slice. A repository-level repair task can reward locating a bug and passing a test suite. It may say little about visual quality, accessibility or whether a product team can maintain the result six months later. Conversely, a model that produces an attractive interface in one demonstration might struggle with a large, unfamiliar codebase.

Developers also change model outcomes through their tools. One team may give an agent a terminal, a repository map, test access and permission to make several attempts. Another may only paste a prompt into a chat window. The same model can behave differently under those conditions. For this reason, readers should treat Google’s chart as a starting point for evaluation rather than a forecast of success on their own engineering backlog.

The phrase “benchmaxxing” often appears in debate over model launches. It suggests optimising for public tests while neglecting other work. There is no evidence in the sources reviewed here that Google deliberately trained Argon to game a specific coding evaluation. The narrower, defensible concern is measurement: a benchmark suite may fail to capture the workflow a company cares about. That concern applies across vendors, and it is best addressed by a private, representative trial rather than an allegation about motive.

A practical model evaluation sequenceSelect real tasks, fix tools and permissions, record accuracy and effort, require human review, and decide deployment scope.From benchmark headline to deployment decisionReal tasksOwn repositoriesSame toolsSame permissionsMeasureErrors and timeReviewHuman approvalDecideLimited rolloutUse one task set for every model; log failures as well as successes.

Access remains limited, so comparisons remain provisional

Google said the first recipients are selected cyber defenders under its Fairwind programme. The company plans later access for developers, enterprises and consumers, beginning with paid API customers and Google AI Ultra subscribers, but it did not set a firm general release date in the announcement. That limits how much outside teams can independently test today. Our earlier coverage of the cyber-defender rollout explains the access boundary and the security rationale in more detail.

Google published an introductory price of $2 per million input tokens and $10 per million output tokens, with later standard prices of $4 and $20 respectively, according to its launch note. These are announced token rates, not a quotation for a complete engineering project. A real pilot also consumes tool calls, developer review, testing time and the cost of correcting mistakes. Procurement teams should verify availability and terms before building a forecast around the introductory rate.

A limited release can be sensible for a model with advanced cybersecurity capabilities. It also means that public commentary may be shaped disproportionately by vendor tests, early testers and confidential accounts. Those are useful signals, but they are not a substitute for a broad independent record. The gap should narrow as access expands and researchers publish repeatable evaluations. Until then, strong claims in either direction deserve a source label.

A useful test for Indian software teams

An Indian startup or services company considering Gemini 4 coding assistance can build a small, safe comparison now, even if it cannot yet run Argon itself. First, define the tasks: a bug from a real repository, a new feature with acceptance criteria, a front-end accessibility fix and a test-writing assignment. Keep the task set private. Then document the allowed context, tools, time limit and human review rules. Those settings should be identical when multiple models are eventually compared.

Record more than whether a test passed. Measure the developer time needed to understand and repair the output, the number of attempts, security issues, regressions and documentation quality. A model that finishes quickly but creates fragile code may cost more overall. A model that performs well on a narrow benchmark may still be useful for a particular team if its own workflow matches the benchmark. The evaluation should answer the buyer’s question, not merely reproduce a leaderboard.

Data handling is another gate. Internal repositories can include customer information, credentials and unpublished business plans. A pilot should begin with a non-production environment and approved data. The team should decide who can grant model access, which actions require review and how changes are rolled back. This is a general deployment principle, not a claim that Argon has a specific security defect. Our report on an AI sandbox incident illustrates why tool permissions deserve as much attention as model quality.

When Argon becomes available to a broader audience, the same evaluation framework can be used to compare it with alternatives. Avoid changing prompts, permissions or task difficulty for one vendor. Keep failures in the results. Report cost alongside quality. If a company needs dependable front-end output, include visual inspection and accessibility tests; a unit-test pass alone will not answer that need. The practical question is whether the tool improves reviewed engineering work within an acceptable budget and risk boundary.

What remains unproven

As of 2 October 2026, Google’s public chart shows both wins and losses for Argon on coding tests. Bloomberg’s reporting adds attributed concerns from some internal users, and Google disputes a broad negative reading. The public material reviewed here does not establish a general failure rate on production software work, nor does it prove that internal criticism represents all users. Likewise, vendor-reported benchmark leadership on selected rows is not a guarantee of a better customer outcome.

That leaves a straightforward editorial conclusion. Gemini 4 coding capability is promising and contested at the same time, depending on the task and the source of evidence. The next meaningful milestones are broader access, independently repeatable tests and the results of controlled pilots on real repositories. A responsible buyer can prepare that evaluation now without declaring either a breakthrough or a disappointment in advance.

Frequently asked questions

Did Google say Gemini 4 Argon is weak at coding?

No. Google disputes the claim that its new model underperforms broadly in coding. Its own published table shows strong results on some coding benchmarks and lower results than listed peers on others. Bloomberg separately reported scepticism from some people with direct access about certain practical tasks.

Can developers in India use Argon today?

Google announced limited initial access for selected cyber defenders and said broader availability would follow. Its 30 September launch announcement did not provide a firm date for general developer access. Buyers should verify current eligibility directly with Google rather than assume an existing Gemini subscription includes the model.

Which benchmark should a company trust?

No single public score should decide procurement. Use benchmarks to select candidates, then test each one on the same private tasks, tool access, review process and budget. Google’s four coding rows already show why the ranking changes with the test.

What does the introductory price cover?

Google announced token prices of $2 per million input tokens and $10 per million output tokens for an introductory period. A company’s full cost also includes supporting tools, repeated attempts, human review and any contractual terms. The listed rate does not by itself establish total project cost.

Sources and method

Primary: Google’s announcement and DeepMind’s performance table. Independent original publisher coverage: Bloomberg, Axios, Ars Technica and VentureBeat. Bloomberg is the original source for the anonymous employee accounts; follow-on summaries are not counted as additional witnesses. Benchmarks and internal productivity examples are attributed to Google.

Frequently Asked Questions

When did Google announce Gemini 4 Argon?

Google announced Gemini 4 Argon on 30 September 2026, with initial access limited to selected cyber defenders.

How does Gemini 4 score on coding benchmarks?

Google reports 77.9% on DeepSWE v1.1, above the peers in its table, but 55.0% on FrontierSWE v2, below three listed rivals.

Is Gemini 4 bad at coding?

Bloomberg cited people with access who questioned some practical coding performance, but Google rejected a broad underperformance label. A supervised trial on your own codebase remains the best test, just as Anthropic’s GLM-5.3 cyber tests show limits of narrow benchmarks.

Get the day’s top stories in your inbox

One concise email. No spam, unsubscribe anytime.