top of page

llm-gemini 0.36 Adds Carbon as Google Expands Paid Access

19 minutes ago
10 min read
Illustrative photograph. Photo: AS Photography / Pexels.

Simon Willison released llm-gemini 0.36 on November 10, adding Gemini 4 Carbon support as Google expanded access beyond its earlier cybersecurity preview. Developers gained another way to evaluate the model, although broader availability still comes with account and product restrictions. The immediate change concerns access and integration, rather than independent proof of better results.

That distinction matters because a model announcement and a working developer integration answer different questions. Google describes the intended capabilities; a plugin makes the model available inside an existing workflow. Neither establishes whether additional reasoning will justify its operational demands on a particular project.

Google also keeps Gemini 4 Argon in the picture, positioning Carbon for more demanding work. The central comparison therefore starts inside Google's own lineup: deeper investigation against the time and resources required to produce an acceptable result. Willison's visual experiments offer an accessible starting point, while leaving repeatability, maintenance, and production suitability open.

What llm-gemini 0.36 Actually Changes

The release gives developers a familiar entry point for Carbon, making practical evaluation easier to separate from launch messaging.

The Carbon release notes identify the addition directly: Gemini 4 Carbon becomes selectable through llm -m gemini-4-carbon. This is an integration release for Willison's LLM software, a command-line tool that lets users interact with language models through text commands.

The distinction between the plugin and the underlying model deserves attention. Updating a client changes what that software can request. Access to the requested service still depends on the provider's requirements, so installing an integration does not automatically establish account eligibility.

For developers already using LLM, the significance is continuity. An evaluation can begin from a familiar interface rather than an entirely different application. That reduces an avoidable source of confusion when comparing model responses: changes in the surrounding software.

A familiar interface also makes differences easier to investigate. If a result changes after switching models, the evaluator has fewer interface changes to account for. This does not eliminate every variable, but it creates a more useful starting point than comparing unrelated demonstrations.

Consider a developer investigating a difficult application bug. The relevant question is whether Carbon identifies the cause and proposes a defensible correction. A polished explanation alone would leave the central issue unresolved, especially if the suggested patch creates another failure.

The release makes that kind of question easier to ask within an established workflow. Its importance comes from connecting an announced capability to an interface people can inspect and incorporate into their own experiments. The connection is practical, even before anyone establishes a performance advantage.

It also creates a distinction between support and coverage. A client can support selecting a model without reproducing every feature available in the provider's applications. Evaluators should avoid assuming that an application demonstration describes the experience of every API client.

The most defensible reading of this release is therefore specific. Carbon has entered another developer access path. Claims about better engineering outcomes require evidence collected after that connection works, using tasks whose success conditions extend beyond producing a plausible response.

That is a useful development without requiring a sweeping verdict. Small integration releases can matter because they let more people interrogate large claims. The value emerges through the evaluations they enable, rather than through the version number itself.

Broader Gemini Carbon Access Still Has Boundaries

Expanded availability creates more evaluation opportunities, but eligibility remains part of the product experience.

Google's Carbon access announcement describes paid API access for developers and enterprises. It also describes a rollout to Google AI Ultra subscribers across selected products, including Antigravity, the Gemini app, and AI Mode.

The announcement follows an earlier release to selected cybersecurity defenders through Google's Fairwind Program. These are different access checkpoints. Treating the November expansion as the model's first appearance would erase the distinction between restricted preview access and the broader commercial rollout.

Equally, broader access should not become shorthand for universal free availability. A product announcement can cover several audiences while assigning different conditions to each. Developers, business customers, and individual subscribers do not necessarily enter through the same route.

For a team evaluating Carbon, that creates a practical procurement question before the performance question. Can the intended users access it through the intended application? An experiment conducted through one person's account does not establish availability across an organization.

Access conditions also affect whether an evaluation is reproducible. If participants use different products, their available tools and permissions can differ. A model comparison becomes harder to interpret when those differences remain undocumented.

The distinction matters particularly when a demonstration circulates without its surrounding setup. Viewers see the finished output, but the account type, interface, and enabled tools may disappear from the discussion. Those details help explain what another user can reasonably attempt.

A useful organizational response is to describe access precisely in any internal assessment. The assessment should identify the product used and the permissions granted. That gives colleagues enough context to distinguish a model limitation from an unavailable feature or account restriction.

There is also a communication issue for product teams considering customer-facing features. Internal access does not automatically establish the conditions under which a feature can ship. Any commitment needs to match the actual service route and terms supporting that implementation.

This does not diminish the expansion. It defines what changed and where further confirmation belongs. More eligible users can investigate Carbon, while specific deployment questions still require answers at the account and product level.

For readers following the announcement, the key development is a widening path to evaluation. Its practical reach should be judged through confirmed availability in the relevant environment. Broad wording alone cannot substitute for that confirmation.

Gemini Carbon vs Argon Is a Resource Decision

The meaningful comparison concerns accepted work, including the effort required to check and repair it.

Google positions Argon as suitable for most software engineering and knowledge work, while describing Carbon as a choice for harder problems. That framing leaves an important decision with users: which tasks justify a more intensive reasoning process?

A task's difficulty cannot be inferred solely from the length of its prompt. A brief bug report can conceal a difficult interaction between components. A lengthy document request can involve straightforward extraction, provided the source material is clear.

The evaluation unit therefore needs to be the completed task. For a coding assignment, completion includes behavior that survives appropriate checks. For a research assignment, it includes claims that remain traceable to evidence after a reviewer examines them.

This changes how apparent efficiency should be interpreted. A quick response that requires extensive correction consumes human attention later. A slower response can still be useful if it reduces that correction burden, although the reduction must be observed.

The reverse deserves equal attention. Extended reasoning does not guarantee that the model is investigating the right issue. A lengthy answer can elaborate on a mistaken assumption, making review more demanding rather than less.

Gemini Carbon vs Argon comparisons should therefore preserve the conditions that matter to the actual decision. The models need comparable source material and a common definition of success. Otherwise, differences in preparation can masquerade as differences in model capability.

For an engineering team, an illustrative comparison would use a real defect with a known reproduction path. Reviewers would examine whether each proposed change resolves that defect. They would also consider whether the change introduces unrelated modifications that increase maintenance work.

For a knowledge worker, the equivalent assignment might involve reconciling conflicting project documents. A satisfactory result would identify the conflict and support its interpretation. Confidently merging incompatible statements would fail the assignment, regardless of how readable the response appeared.

These examples describe evaluation criteria, rather than results established for either model. Their purpose is to make the decision concrete. Without explicit criteria, the evaluator risks rewarding whichever answer feels more impressive during an initial reading.

The organizational pressure falls on whoever chooses the default model. Selecting Carbon for every request requires justification, just as excluding it from difficult work requires justification. The useful policy connects model choice to observed task requirements.

A selective deployment would strengthen the case for Carbon if demanding assignments show a consistent benefit. Little improvement on routine work would support keeping a less intensive option available. Either finding would be more actionable than a general declaration that one model is best.

Thinking Controls Change the Evaluation

A configurable reasoning setting creates another variable that comparisons need to disclose.

The plugin documentation lists low, medium, and high thinking levels for Carbon. A thinking level controls how much reasoning the model performs before responding, although the consequences still need evaluation on the task at hand.

That setting matters because a model name alone provides an incomplete description of an experiment. If evaluators use different reasoning levels, they are testing different configurations. A useful comparison records the setting alongside the prompt and available tools.

This is especially relevant when someone reports a disappointing result. The result remains valid evidence about that attempt, but its scope depends on the conditions. Without those conditions, readers cannot tell what they should expect from a comparable request.

The same standard applies to favorable examples. An impressive output from a selected configuration does not establish that every configuration produces equivalent work. Presenting the setting helps readers interpret the result without inflating its significance.

There is a deeper issue concerning what additional reasoning is supposed to accomplish. An evaluator should identify the expected benefit before examining the answer. Otherwise, any increase in detail can appear to confirm that the model thought more effectively.

For a debugging assignment, the expected benefit might be a more accurate causal explanation. For a document review, it might be better handling of contradictory evidence. Those are observable qualities, provided the reviewer has enough information to judge them.

Response length offers a weaker signal. More text can contain useful investigation, repetition, or unsupported explanation. Its value depends on whether it advances the assignment and reduces uncertainty for the person responsible for the result.

Thinking controls also complicate default choices. A default setting becomes part of the experience for anyone who does not change it. Teams should recognize that their conclusions might reflect that default rather than the full range of available configurations.

For readers searching for Gemini Carbon explained, this is the operational distinction to retain. The model offers configurable reasoning behavior. The configuration belongs in the evidence whenever someone uses an experiment to support a broader claim.

The release supports investigating that distinction through a documented client interface. It does not settle which setting belongs in every workflow. That decision depends on measured quality, tolerable delay, and the amount of human review the task requires.

The Pelicans Show Possibilities, With Limited Evidence

A visible creative result can reveal useful behavior while leaving reliability and comparative performance unresolved.

In his Carbon coding experiments, Willison published pelican examples across thinking settings. He also described an Antigravity experiment involving a three-dimensional animation of Yonkers becoming a large mech and moving through New York City.

The animation request used Three.js, a JavaScript library for creating browser-based three-dimensional graphics. Willison reported an additional prompt to obtain a standalone HTML version without external dependencies. He also referenced an earlier attempt with Claude Fable 5.2.

These details make the example useful as a report of an interaction. The request is concrete, the output is inspectable, and the follow-up reveals an additional requirement. Readers can understand what interested the evaluator without treating his reaction as a formal measurement.

The follow-up is particularly informative because delivery format affects usability. A generated scene can satisfy a visual idea while still needing packaging work. Asking for a standalone result introduces a practical requirement beyond the initial creative request.

For an evaluator, that suggests a distinction between generating an artifact and delivering it in the needed form. Both belong in an assessment when the intended user depends on the final package. Ignoring the latter can understate the work required after generation.

The pelican examples have a different evidentiary role. They make outputs easy to inspect and discuss. They do not establish how consistently the model handles unrelated assignments, especially those involving hidden constraints or correctness that requires specialist review.

The Claude reference also needs careful handling. Reusing an idea offers a recognizable point of comparison, but a model ranking requires more controlled evidence. Differences in prompts, tools, follow-up requests, and selection of outputs can affect the apparent result.

There is no need to dismiss informal experiments because they lack those controls. They can expose behavior worth investigating and make technical releases understandable. The limitation concerns the conclusions drawn from them.

A useful demonstration can motivate a focused question for subsequent testing. Does the generated project remain understandable after someone edits it? Does it preserve requirements when asked to change one component? These questions extend naturally from a creative artifact into engineering concerns.

For knowledge work, the parallel is a report that reads well during a presentation. Its appearance gives viewers something to assess immediately. Its evidence, handling of uncertainty, and suitability for a consequential decision require closer examination.

The appropriate conclusion is bounded but useful. Willison's examples supply concrete material for discussion and further evaluation. They leave open the broader question of whether Carbon improves repeatable outcomes under the conditions that matter to a particular user.

What to Watch After the Expanded Rollout

The next useful evidence concerns confirmed access, repeatable task outcomes, and deployment limitations.

The first signal is whether access becomes consistently available through the products users intend to adopt. A rollout announcement establishes direction. Confirmed access in the relevant environment establishes whether an evaluation or deployment can actually proceed.

That signal would strengthen the practical significance of the expansion if teams can reproduce their intended setup. Persistent uncertainty around eligibility would narrow its immediate usefulness. This is an adoption question, separate from whether the model produces a strong answer once reached.

The second signal is independent evaluation that reports complete task outcomes. The most useful evidence will disclose prompts, tools, reasoning settings, and acceptance criteria. Those details let readers decide whether an observed benefit transfers to their own work.

Repeated success on difficult assignments would support the argument for allocating more reasoning resources selectively. Results that require substantial repair would weaken the case for those assignments. Either outcome should be described with the scope of the experiment intact.

The third signal is how documented limitations appear in ordinary use. Google's Carbon model card identifies hallucinations, occasional slowness or timeouts, and increased token use at higher thinking levels. Tokens are units of text processed or generated by a model.

The card also distinguishes specialized cybersecurity capabilities available to verified defenders. That distinction matters when interpreting claims associated with the earlier restricted release. Readers should avoid assuming that every capability discussed in that context transfers unchanged to public access.

These limitations give evaluators concrete issues to observe without predetermining the result. A timeout matters differently in an exploratory session than in a scheduled business process. An unsupported statement matters differently when a reviewer catches it before delivery than after someone acts on it.

The evaluation should preserve those consequences. Recording only the final successful output can hide the attempts and interventions required to reach it. Recording only the initial failure can obscure whether a reasonable follow-up resolves the problem.

For teams collecting evidence, a searchable knowledge base can keep prompts, source documents, review notes, and accepted changes connected. That supports a more traceable decision when colleagues revisit why a model was chosen.

Together, these signals would clarify where the release has practical value. Availability determines who can investigate. Repeatable outcomes determine which tasks benefit. Documented limitations help define where additional review or a different approach remains necessary.

Before making llm-gemini 0.36 part of a regular workflow, choose a task whose successful outcome you can recognize and defend. Preserve the original requirements, record the model configuration, and examine the result with the same care you would apply to another contributor's work. Include the time spent correcting mistakes and checking assumptions. Compare the accepted result with your existing approach, including Argon where relevant, rather than comparing presentation quality alone. Keep the examples that fail alongside those that succeed. The release gives eligible developers another route to investigate Carbon's behavior; use that access to build an evidence trail for a specific decision. Which demanding task would make that investigation worthwhile for your team?

Give every agent the context to do better work

Connect your agents to the knowledge, decisions, and history already organized in remio.

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