Google Gemini

We fed Gemini 2.5 Pro a complete Next.js codebase, approximately 600,000 tokens, the equivalent of 450,000 words, and asked it to trace a bug spanning four separate modules. It identified the root cause in a single pass. No chunking, no file-by-file uploads, no retrieval-augmented workarounds. Just the entire codebase loaded at once, and one answer that connected the dots across all four modules simultaneously.

That test tells you where Gemini earns its place in 2026. The million-token context window is not a spec sheet number, it is a practical advantage that no competing consumer model currently matches. But the same tool that impresses you on a Tuesday will, on a Thursday, ignore the instructions you gave it two prompts ago. That inconsistency is the real subject of this review.

What Gemini ships as of September 2026

Gemini has gone through more naming changes than almost any AI product in the market. What launched as Bard in 2023 became Gemini in February 2024. "Gemini Advanced" became Google One AI Premium, then Google AI Pro at Google I/O in May 2025, when a new $249.99/month Ultra tier also launched. As of September 2026: Free Gemini app; Google AI Plus starting $7.99/mo (400 GB); Google AI Pro commonly $19.99/mo; Google AI Ultra split into Ultra 5x $100/mo and Ultra 20x $200/mo (I/O 2026 reduced the prior $249.99 top tier).

API/DeepMind Flash ladder as of Sep 2026: Gemini 3.7 Flash (latest Flash for agentic/coding) and Gemini 3.6 Flash (prior Flash), plus 3.5 Flash / Flash-Lite and 3.1 Pro Preview. Image generation is officially Nano Banana (Nano Banana 2 / 2 Lite / Pro). The Pro subscription unlocks a 1 million-token context window, Deep Research (autonomous browsing of 100+ sources with export to Google Docs), and integration across Gmail, Drive, Docs, Sheets, and Slides. If you live in Google Workspace, that integration changes the comparison against standalone AI tools.

What Gemini is actually good at

Long-context analysis is the headline capability. The codebase test we ran is representative of how this plays out in practice. Gemini 2.5 Pro achieves 100% recall up to 530,000 tokens and 99.7% recall at one million tokens, the largest reliable context window of any major consumer AI model. For comparison, GPT-4o tops out at 128,000 tokens and Claude 3.7 at 200,000. Loading a large codebase, a lengthy legal document, or an entire product backlog into Gemini without splitting it up is a genuine workflow advantage, not a marketing talking point.

"I've been having some very impressive results from Gemini 2.5 Pro for complex coding tasks.". Hacker News commenter, March 2025

A follow-up in the same thread: a developer described Gemini one-shotting a fully functional LISP interpreter, not a skeleton, a working implementation. Long context combined with strong multi-step reasoning is where Gemini competes at the top of the market.

Deep Research has a UX edge that matters. Before executing a research task, Gemini presents its research plan and asks whether you want to edit it. No other major AI tool offers this step. You can steer the research toward specific sources, away from certain angles, or toward a depth level that matches what you need. Reports include source mapping so you can audit the reasoning behind any conclusion, and export directly to Google Docs in one click.

The reasoning transparency in thinking mode is unusual. Gemini 2.5 Pro will sometimes reason through a task visibly, then stop to tell you the task is not feasible in one shot rather than producing a confident wrong answer. One developer described it as "a new type of refusal" in a positive sense, the model declining to fake competence.

The Deep Research plus Audio Overview combo is a non-obvious productivity win. Run a Deep Research report, then convert it to a podcast-style audio conversation, two AI voices, deliberate pauses and banter to feel conversational rather than read aloud. Users listen during commutes. One click from the research output inside Gemini.

Google Workspace integration is a real differentiator for teams that live in Gmail, Docs, or Drive. Since January 2025, Gemini AI features are included in all Google Workspace Business and Enterprise plans at no extra cost, replacing a previous $18/user/month add-on. For teams already paying for Workspace, that changes the comparison against standalone AI subscriptions.

Where Gemini breaks, the failure modes users keep hitting

The single most consistent complaint across Reddit, Hacker News, developer forums, and review sites is not any specific bug. It is a pattern of inconsistency and regression: Gemini works well on one task, then fails on the same task later. Instructions followed in one session are ignored in the next. A model version that developers learn to rely on gets superseded by a version that breaks their workflows.

Gemini 3.0 becoming a documented "major downgrade" from 2.5 Pro is the sharpest recent instance of this. A Google AI Developers Forum thread gathered multiple developers reporting that the newer model failed at tasks the older one handled reliably.

"The model doesn't stick to instructions at all, it completely falls apart, often by the 2nd or 3rd generation.". Google AI Developers Forum, 2025

A separate developer in the same thread added that Gemini was "refactoring and removing parts of my codebase that I explicitly told it not to remove," and noted complete loss of work when context overflow caused conversation history to be deleted on page reload.

API reliability is a documented problem. A Hacker News thread gathered developer accounts of responses stopping mid-sentence, not because of token limits, but what appeared to be a completion-signaling bug. A separate report noted that Gemini cannot perform tool calling and forced JSON output simultaneously: one developer described retrying half their API requests because of consistent failures, with a support ticket open for months without resolution.

Hallucinations in cross-app actions are well-documented. When Gemini operates across Google apps, calendar, Docs, Drive, its error rate is higher than a $19.99/month subscription should produce. One widely-shared cancellation piece described Gemini creating non-existent events in a calendar from fabricated source data. Readers chimed in: "The majority of things I have asked it to do it failed."

Deep Research output quality is uneven. The interface is excellent, planning step, source mapping, Docs export, but the content can disappoint. A head-to-head test scored Gemini's output 2.5/5 on quality versus ChatGPT's 4.5/5, with Gemini reports being "too high level to be immediately usable" and defaulting to mainstream sources rather than surfacing novel insights.

Gemini vs. ChatGPT vs. Claude

ChatGPT (GPT-4o / GPT-5) is the most common point of comparison. Reddit consensus is consistent: ChatGPT feels like "a smart friend" while Gemini reads as corporate and formal. For creative writing, marketing copy, or any task where personality matters, ChatGPT has a durable edge. Its 128,000-token context limit is a structural disadvantage for large document work, but for most daily tasks the tone gap is the deciding factor.

Claude (Sonnet / Opus) is the preference among developers who need reliable instruction-following across long sessions. Hacker News developers noted that Claude was "easier to steer toward the desired implementation" compared to Gemini, which would deviate from agreed plans mid-session. Claude 3.7 Sonnet scores 70.3% on SWE-Bench versus Gemini 2.5 Pro at 63.8%, a measurable coding gap. Claude's 200,000-token context window is smaller than Gemini's million-token ceiling, but for most codebases it is sufficient, and session consistency often matters more than peak context size.

The decision tree: Gemini when context length or Workspace integration is the bottleneck, ChatGPT for creative work with personality, Claude for sustained complex instruction-following.

Is the paid tier worth it?

The free tier is more capable than most users realize. Gemini 3 Flash with daily limits on Pro and Thinking models, 32,000-token context, and up to 20 image generations per day is a functional starting point. The limitation that matters is the context window: 32,000 tokens versus one million tokens on Pro is a meaningful difference for any use case involving large documents.

Google AI Pro at $19.99/month is where the substantial capability jump lives. The million-token context window alone justifies the cost for developers working with large codebases or research collections. Add Deep Research, access to Gemini 2.5 Pro and Gemini 3, the Jules coding agent, enhanced NotebookLM with 5x audio overviews, and Veo 3.1 video generation, and the tier is genuinely dense with features.

The concern with Pro is reliability. Paying $19.99/month and regularly hitting errors, disappearing chat history, or responses that ignore your latest message is a frustration that appears in reviews consistently enough to take seriously. The Android Police cancellation article, widely shared, written by a mainstream user rather than a developer, framed it plainly: "I'm sitting here wondering when, exactly, is Google going to prove that I need to be paying for Gemini Advanced."

Google AI Ultra 5x at $100/mo (I/O 2026 blog): from 20 TB, 5× Pro usage, 10,000 Flow credits, Deep Think / higher Omni, YouTube Premium, $40 Cloud credits. Ultra 20x at $200/mo (reduced from $250/$249.99 era): 30 TB, 20× Pro usage, 25,000 Flow credits, Project Genie, $100 Cloud credits. I/O 2026 also moved app limits toward compute-used pools refreshing ~every five hours until a weekly cap.

Best use cases (and when to skip it)

Use Gemini when: you have large files or codebases exceeding 200,000 tokens and need to analyze them without chunking; you want a research tool that lets you edit the plan before it executes; you live in Google Workspace and want AI inside Gmail and Docs rather than a separate tab; or you want to convert research into audio you can listen to passively.

Skip Gemini when: you need reliable instruction-following across many turns. Claude is more consistent; you are building on the API and need predictable output format and completion behavior, the truncation and JSON/tool-call bugs are real enough to architect around; or you want creative work with warmth and personality. ChatGPT still leads there.

Getting started with Gemini

The fastest test of Gemini's context advantage: load a large document, a long PDF, a multi-file codebase as plain text, a lengthy research corpus, and ask questions that require connecting information across sections. This is the use case where Gemini has no consumer competition at its price point.

For Deep Research, start with a complex question you would normally spend an hour investigating manually. When Gemini presents its research plan before executing, edit it: specify source types, exclude angles you do not need, set scope. Most new users skip this step and get generic output as a result. Editing the plan produces measurably better reports.

The free tier's daily access to Pro and Thinking models gives a reasonable sample before committing. The context window gap, 32,000 tokens free versus one million on Pro, is the ceiling you will hit first if large documents are part of your work. As-of / last verified pricing & model IDs: 2 September 2026.