top of page

llm-gemini 0.38 Adds Gemini 4 Neon

23 minutes ago
10 min read
Image from the source report. Source: simonwillison.net.

Simon Willison released llm-gemini 0.38 on November 18, adding Gemini 4 Neon support while Google's expanding model lineup complicates selection. The immediate change is concrete: developers using his LLM application can select Neon with llm -m gemini-4-neon. The harder question is whether another accessible model makes their decisions easier.

This is a plugin release, not a separate Google model announcement. A plugin is an extension that adds capabilities to an existing application. Here, the addition puts Neon inside a familiar development workflow, where users can assess it against alternatives.

The central tension is access versus judgment. Adding a model identifier lowers the effort required to try Neon, but it does not establish where Neon belongs in production. Google's Carbon, Hydrogen, and Argon models remain part of the comparison, and their names alone cannot settle the choice.

What llm-gemini 0.38 Actually Changes

The release adds a specific route to Neon, rather than evidence that developers should replace their existing model.

The project's version 0.38 release identifies Gemini 4 Neon support and gives the command used to select it. That is the verified event at the center of this update.

The distinction matters because announcements often compress several separate developments into a single story. A provider introduces a model, an integration exposes it, and users later establish whether it improves their work. Those developments have different evidence requirements.

For this release, the relevant change happens at the integration layer, the software connecting an application to a model provider. Existing LLM users gain another named option without adopting a completely different application.

That makes experimentation more convenient. It does not mean every feature advertised by the model provider automatically appears through the plugin. Support for a model and support for every possible interaction are separate questions.

An application might successfully send a text prompt while requiring additional handling for attachments, external tools, or specialized settings. Readers should therefore interpret the release according to its stated scope, rather than infer a complete feature checklist.

The naming also deserves attention. The number 0.38 identifies the llm-gemini plugin release. It is not the version number of Google's underlying model or a claim about the model's relative capability.

That separation helps explain why this comparatively small release deserves coverage. Much of a model's practical usefulness depends on whether it reaches tools people already use. Integration can turn an announcement into something developers can actually evaluate.

Consider an engineer investigating inconsistent answers in an internal document assistant. Access through an established application makes it easier to include Neon in an investigation without rebuilding the surrounding workflow. This is an illustrative scenario, not a reported customer deployment.

The engineer still needs to preserve the documents, instructions, and acceptance criteria used in earlier tests. Otherwise, a better-looking answer might reflect a changed prompt rather than a better model.

That is the immediate opportunity created by the release: a simpler entry point for comparison. The result of that comparison remains open, and the release note supplies no independent performance verdict.

Readers should also keep the publication timeline straight. November 18 is the date attached to the plugin release. Reading about it on November 20 does not make it a new model launch on that later date.

The useful news judgment is therefore narrow but meaningful. Neon has become accessible through this particular integration, giving existing users another candidate to test against their actual work.

Easier Access Puts More Pressure on Model Selection

Developers gain another option, while whoever owns reliability must decide whether that option earns a place in the workflow.

The LLM application documentation describes a command-line application and Python library for interacting with models from multiple providers. A command-line application accepts instructions through a terminal rather than a graphical interface.

That shared environment provides context for the update. Google's Gemini is one model family available to LLM users; OpenAI and Anthropic models are also represented in the broader application.

This does not make those providers interchangeable. It means developers can approach different systems through a common application while still accounting for differences in behavior, settings, and availability.

The comparison pressure falls first on developers responsible for choosing defaults. A default determines which model receives work when users make no explicit selection. Changing it can affect every downstream task that relies on that choice.

For those developers, a new option creates an evaluation obligation rather than an automatic migration. They need a reason to move beyond curiosity, especially when the current model already meets requirements.

Product managers face a related question. A new model's appeal must translate into something users experience, such as more accurate answers or less waiting. A model name on a settings screen does not establish either outcome.

Technical buyers have a different burden. They must distinguish a useful experiment from a dependable operating choice. The relevant evidence includes repeatable behavior and applicable commercial terms, not just a successful demonstration.

These pressures arise even when the integration itself is straightforward. Ease of access can accelerate experimentation faster than a team can interpret the results. More trials do not necessarily produce clearer decisions.

An illustrative support workflow shows the issue. Suppose a model drafts answers from approved product documents. A candidate model might write more naturally while overlooking a qualification that the existing model consistently preserves.

A casual reviewer might favor the smoother response. A support lead might reject it because the missing qualification changes the customer's understanding. Both reviewers saw the same output, but they applied different standards.

That disagreement is why selection criteria should precede a migration decision. The organization needs to know whether correctness, speed, consistency, or editorial quality determines success for that particular task.

This is analysis of the decision created by the release, not evidence that Neon exhibits any particular failure. The public release establishes availability through the plugin. It does not establish the consequences for an individual organization's workload.

The pressure also reaches integration maintainers. Users naturally want familiar controls to behave predictably when a new model arrives. Maintaining that expectation requires clear boundaries around what the integration supports.

The practical consequence is a division of responsibility. The plugin makes the option reachable; the provider describes the model; the adopting team decides whether the combination meets its requirements.

That division is easy to overlook when model announcements and integration releases arrive close together. Keeping it visible prevents a small compatibility update from becoming an unsupported claim about production readiness.

Google's Model Lineup Resists a Simple Ranking

Neon joins a family of choices, and a position in that family is not a universal recommendation.

Google's Gemini model catalog lists Carbon, Neon, Hydrogen, and Argon. It describes them with different emphases, including reasoning, task completion, and efficiency.

Those descriptions establish how Google presents the lineup. They do not independently establish which model will deliver the best result for a particular application.

This distinction becomes especially important when readers encounter simplified graphics or short launch summaries. A compact presentation can help introduce a product family while leaving out information needed for an operational decision.

For example, a visual ordering might imply a straightforward progression from lighter work to harder work. Actual model selection also depends on what constitutes success and which failures the application can tolerate.

A document extraction task and an open-ended research task do not ask the same question of a model. Extraction requires faithful handling of supplied material. Research assistance may additionally require careful uncertainty and source evaluation.

Even within the same application, different tasks can justify different choices. Drafting an initial response and checking a final answer create distinct responsibilities. There is no requirement that a single model perform both equally well.

The available evidence also argues against treating an abbreviated lineup as proof that another model has disappeared. Argon remains listed in Google's catalog, alongside the other Gemini 4 models.

That is a useful correction to any interpretation based solely on omission from a promotional presentation. An omitted name and a discontinued service are different claims, requiring different evidence.

Commercial comparisons introduce another complication. Google's API billing terms specify a scheduled change for Carbon beginning December 1, 2026. An API is the interface through which software sends requests to a service.

That date makes the comparison time-sensitive. A conclusion drawn under one set of terms should not automatically carry into a later operating period. Readers need the effective date alongside the model identity.

The broader inference is straightforward: a product hierarchy and an economic hierarchy need not match. Promotional positioning describes intended roles, while actual operating economics depend on the applicable terms and workload.

This article does not reproduce price figures or declare a universally cheapest model. Such a declaration would require a defined service, date, request pattern, and measurement of the work actually completed.

For buyers, that means an apparently simple comparison can conceal different assumptions. A brief answer, a lengthy analysis, and a task requiring repeated attempts can produce different operating outcomes.

The same caution applies when comparing Google with OpenAI or Anthropic. A shared application can make alternatives easier to reach, but a shared interface does not erase differences between the services.

The relevant competitive pressure is therefore practical. Each model must justify its role against the alternatives already available to the user, rather than benefit from its position in a launch narrative.

Neon's arrival expands that contest. The plugin release makes participation easier, but the user's workload still determines which competitor has earned the assignment.

Model Support Is Not a Performance Verdict

The strongest skeptical reading separates successful integration from successful work.

Google's Neon model specifications describe its supported inputs, text output, and available capabilities. Google positions Neon for software engineering, knowledge work, and scientific workflows.

Those are provider descriptions. They give developers a basis for deciding what to investigate, but they do not replace independent testing in a specific application.

The plugin's Gemini integration documents model selection and access to the Gemini family. That establishes the integration's purpose without proving comparative accuracy, adoption, or business outcomes.

These boundaries matter because an apparently successful response can satisfy a superficial test. A model can follow the requested format while introducing an unsupported detail. It can also produce correct content too slowly for the intended interaction.

Neither possibility is a demonstrated Neon weakness here. They illustrate why receiving an answer is an incomplete standard for evaluating a model.

A credible comparison needs examples that represent the actual work. Convenient prompts can help reveal basic behavior, but they often omit the difficult cases that determine whether a system remains useful.

For a research assistant, difficult cases might include contradictory documents or questions with no supported answer. For a coding assistant, they might include a change that appears plausible but violates an existing requirement.

These are proposed evaluation categories, not reported benchmark results. Their purpose is to clarify what evidence would support a stronger adoption claim.

Reviewers should also distinguish quality from preference. A response can sound more confident, concise, or polished without becoming more accurate. Presentation matters, but it should not quietly replace the original acceptance criteria.

An equally important uncertainty concerns configuration. If evaluators change instructions or settings while switching models, they are comparing combinations of choices. They cannot confidently attribute the difference to the model alone.

That does not make such comparisons worthless. It changes what they establish. An organization may reasonably compare the best practical configuration of each candidate, provided it records that decision.

The release itself supplies no basis for saying Neon wins such a comparison. It also supplies no basis for saying the added choice is unnecessary. Both conclusions would exceed the evidence.

This is where modest claims become more useful than dramatic ones. The verified integration gives developers an opportunity to investigate a model. The value emerges when the investigation produces a repeatable improvement.

Teams should preserve unsuccessful examples as carefully as favorable ones. A difficult failure can identify a boundary that a collection of attractive responses never reveals.

The same principle applies to later reviews. If an evaluation survives only as someone's recollection that a model seemed better, the organization has little basis for revisiting that judgment.

Keeping technical documents and decisions searchable supports that continuity. A searchable knowledge base can help teams retain the context behind model choices.

That connection does not imply that remio supplies Neon evaluation or supports this plugin. It concerns the surrounding knowledge work: retaining requirements, observations, and decisions so future reviewers can understand them.

The skeptical conclusion is therefore constructive. Treat easy access as the beginning of evidence collection, and reserve performance claims for results that withstand the workload's actual demands.

What Developers Should Watch Next

The next useful evidence concerns integration behavior, dated commercial terms, and repeatable workload results, in that order.

First, watch what happens when users exercise the integration beyond the initial model selection. Clear documentation and reproducible issue reports will help establish the practical boundaries of Neon support.

The relevant question is specific: does the integration reliably support the interactions an adopting team needs? A working basic request is useful evidence, but it cannot answer every compatibility question.

Subsequent fixes would need interpretation rather than an automatic negative verdict. A resolved problem can improve the integration, while repeated unresolved failures would weaken confidence for the affected use case.

This signal strengthens the access argument if users can reproduce expected behavior across their required tasks. It weakens that argument if reaching the model still leaves important workflow requirements unsupported.

Second, watch the December 1 commercial transition identified in Google's documentation. Comparisons intended for ongoing use should reflect the terms applying during that use, rather than preserve assumptions from an earlier period.

The important observation is whether the adopting team's preferred choice changes under the applicable conditions. That requires a defined workload, because a model's commercial position alone cannot determine its overall suitability.

If teams maintain the same preference after updating their assumptions, the original decision gains support. If their preference changes, that would reinforce the importance of dating every comparison.

Neither result would establish a universal winner. It would show how sensitive a particular decision is to the conditions surrounding it.

Third, watch for evaluations that publish enough detail to explain the result. Useful evidence should identify the model, task, instructions, relevant settings, and standards used to judge success.

The strongest results will include difficult examples and failures alongside favorable outputs. That allows readers to decide whether the evaluation resembles their work instead of simply accepting the evaluator's preference.

A convincing advantage on representative tasks would strengthen the case for adopting Neon. An advantage limited to convenient demonstrations would leave the broader case unresolved.

Repeated results also matter more than isolated surprises. A spectacular answer can suggest a promising direction, while consistent performance gives a team something it can plan around.

These signals put the release in perspective. The addition is already meaningful for users who wanted this access path. Its larger significance depends on what becomes possible, dependable, and worthwhile afterward.

For now, llm-gemini 0.38 gives existing LLM users a concrete way to investigate Neon. It does not settle the competition within Google's lineup or across other providers.

That measured conclusion leaves room for a strong outcome. An integration does not need to prove a market shift on release day to become useful infrastructure for subsequent decisions.

Before changing a working setup, identify the task that currently fails and the evidence that would justify a different model. Preserve the instructions, source material, and acceptance criteria used to make that judgment. Compare the candidate under the commercial conditions that will apply when you intend to use it. Keep favorable answers and difficult failures together, so later reviewers can understand the choice. If you are following llm-gemini 0.38, begin with the verified addition and ask what it enables in your own work. The next decision should rest on a repeatable result, rather than the appeal of another model name. Which unresolved task would make Neon worth adopting in your workflow?

Get your work moving in remio

See what needs your attention and take the next step in Workspace.

remio currently supports Windows 10+ (x64) and Macs with Apple silicon.

Your AI Partner at Work
Get more done with remio

Plan. Create. Deliver.
All in one place.

bottom of page