The Slow, Sad Decline of Gemini’s Promise

Gemini once looked like Google’s answer to the whole new chatbot era. ChatGPT had the first-mover advantage, Claude was winning over people who cared about writing and coding, and Perplexity made search feel conversational. Then Gemini started giving people reasons to believe Google could catch up and even lead.

In 2024, Gemini 1.5 Pro arrived with a one-million-token context window for Gemini Advanced. Google was also showing Project Astra as a vision for an assistant that could see, listen, and act in real time. A year later, Google said Gemini 3 Pro topped the LMArena leaderboard and reported strong results on coding and agent benchmarks. Those were ambitious releases. For a while, Gemini felt less like a late entrant and more like the next platform to watch.

That history is what makes the current experience feel so disappointing. This is not a story about Google forgetting how to build models. Google announced Gemini 4 Argon on September 30, then began broader developer and consumer access on October 13. The decline is in the bargain around the model: what you can choose, where you can run it, how much agent work you can do, and how much trust Google asks you to place in its account rules.

A free model picker with less choice

Google’s current Gemini help page says that people without a Google AI plan get “Auto or Flash-Lite” in chat mode. With Auto enabled, Google chooses the model and says that most prompts use Flash-Lite, while prompts needing more reasoning may be routed to Flash or Pro. Turn Auto off and every response uses Flash-Lite. If what you want is a simple, consistent way to choose the stronger model yourself, the free experience does not offer it.

That distinction matters. “The model is available somewhere in the system” is not the same as “I can decide which model handles this task.” Users have less control over quality, and it is harder to know what they are getting before they spend time on a long prompt or a serious coding problem. When the usage limit is reached, Google says Gemini continues the conversation with Flash-Lite. A powerful model that is only available through automatic routing or until a quota runs out feels more like a preview than a dependable tool.

Antigravity puts the agent behind another set of gates

Google’s answer for agentic development is Antigravity: an IDE, command-line client, and agent platform that can work across a project. The idea is good. The access rules are harder to love.

Antigravity’s own plan documentation says rate limits and model availability depend on the Google AI plan. Baseline quotas refresh weekly for users without Pro or Ultra; Pro and Ultra get more generous limits that refresh every five hours until their weekly cap. Google also says quota use depends on how much work the agent does, so a complex task consumes more than a quick edit. That may make sense operationally, but it makes planning a long coding session difficult: the unit developers care about is completed work, while the meter tracks work behind the scenes.

The model menu is tiered too. The current model table puts Gemini 4 Argon behind paid Pro or Ultra access, while several Gemini models are listed for Enterprise only. In a market where developers routinely compare ChatGPT, Claude, Gemini, and other tools by how useful they are inside a real workflow, a plan chart can matter as much as a benchmark score. The best model does not help much if the path to use it is gated, capped, or hard to predict.

Argon is a real contender, with an uneven scorecard

One distinction helps make the comparison fair: Gemini 4 Argon is a model; Antigravity is Google’s agent platform. Argon is now available through the paid Gemini API, in Gemini apps for Pro and Ultra subscribers, and in Antigravity for eligible plans. Google reports a 77.9% score on DeepSWE v1.1 and 51.3% on Zapier’s AutomationBench. Those are strong results, although they are figures published by Google rather than a neutral comparison across every competing model.

For a closer comparison, Vals AI ran a group of current models on Terminal-Bench 4.0 using the same Mini-SWE-agent harness. The benchmark uses one shell tool and 66 long, varied tasks; Vals ran the full set three times and reports the average. Its leaderboard, updated October 16, shows this result:

On this test, Argon lands close to Astra and Fable, trails Opus by about 7.6 points, and scores above GPT-6.1 Sol. Its 0.5-point gap to Fable and 2-point gap to Astra are close; Vals reports an average of three full runs, so those small differences should not be treated as decisive. OpenAI lists GPT-6 Astra and GPT-6.1 Sol as separate models, so they appear separately here too. Anthropic likewise distinguishes the generally available Fable 5.1 from Opus 5.5. The cost column is Vals AI’s measured cost per test at the listed token prices. Google is currently advertising an introductory Argon API price of $2 per million input and $10 per million output tokens through November 30; its stated standard price after that is $4 and $20. The temporary price changes the cost comparison, while the score remains the same.

These numbers do not settle which full coding agent is best. Vals gives each model the same basic shell harness; it does not compare Antigravity with Claude Code or Codex, and Terminal-Bench cannot represent every kind of development work. Still, it is a useful check on the claim that Argon is either an uncontested leader or a weak model. It is a capable frontier model, close to two rivals on this run, behind Opus, and expensive at its announced standard rate. The visible product constraints are therefore a real competitive disadvantage even when the model itself is competitive.

The agentic strategy narrows the developer’s choices

The friction gets sharper when developers want to use a familiar harness. Google’s Antigravity terms explicitly say that third-party software, tools, or services used to access Antigravity—including OpenClaw with Antigravity OAuth—breach the agreement. The terms say this can be grounds to suspend or terminate Antigravity and/or Gemini CLI accounts. Google’s separate Gemini prohibited-use policy says repeated confirmed violations may also lead to restrictions on a person’s Google account.

That is a serious constraint for agentic development, where the harness is part of the tool. Developers want to connect a model to the editor, terminal, browser, scripts, and approval flow they already use. Google can offer a Gemini API with its own terms and billing, but subscription-backed Antigravity access is expressly tied to Google’s own tools. A developer should not have to wonder whether trying a different agent interface could cost them access to a paid coding product.

This was not just hypothetical enforcement. In the Google Gemini CLI repository, users reported losing Antigravity or CLI access after using third-party tools with account authentication. A Google representative acknowledged “recent bans linked to third-party tool usage” and said Google was rolling out an automated unban for affected accounts. The documented consequence concerns Gemini development products and accounts; Google’s broader policy separately describes when a Google account may be restricted. That mix of strict terms, account-level risk language, and a later mass unban is a poor trust story, even when Google reverses mistaken or overly broad enforcement.

Google also announced that consumer Gemini CLI and Gemini Code Assist extension requests would stop on June 18, 2026, directing Pro, Ultra, and free individual users toward Antigravity CLI. Enterprise access and Gemini CLI with paid API keys remain. Consolidating products can focus engineering effort, but for developers it means migrating an established path into the platform Google has chosen. A company that wants to lead agentic coding should make its model easier to integrate, not make the approved route narrower with every product transition.

A frontier model is not enough

Gemini can still compete on raw capability. Google’s own recent releases make that clear. But the agent race is not decided by model quality alone. It is also decided by whether people can choose the model, run it where their work already lives, understand the limits, and trust the rules when they sign in.

That is where Gemini’s promise has dimmed. ChatGPT, Claude, Perplexity, and the open developer ecosystem trained users to expect AI to show up wherever the task is. Google’s strategy keeps pointing in the other direction: take the model, the harness, and the quota as one bundled experience, then ask developers to migrate when the bundle changes.

Gemini is not dead, and the models are not the sad part. The sad part is watching a company with the research, distribution, and infrastructure to set the pace make its agent story feel conditional. If Google wants developers to bet on Gemini again, it needs predictable quotas, clear model access, a dependable appeals process, and an integration policy that gives developers room to build. Agentic AI is a race to get useful work done. A model that cannot travel with the developer’s workflow is already starting from behind.