Google DeepMind's Gemini 1.5 Pro changed what long-context AI could do, and it's already been pulled from active API shelves. The model, announced for early testing on February 15, 2024, and made generally available on May 23, 2024, reached a 2 million-token context window by June 27, 2024. That isn't a vanity stat. It meant a developer could drop an entire codebase, a 300-page contract, and a 90-minute video transcript into a single prompt and still get a coherent answer back. According to Gate.AI's documentation, the model was retired by mid-2026, with version IDs gemini-1.5-pro-001 and gemini-1.5-pro-002 marked off the active list. The retirement is quiet, but the footprint is loud.
The numbers are concrete. Gemini 1.5 Pro used a Mixture-of-Experts architecture and accepted text, code, image, audio, video, and document inputs. Its standard window started at 128,000 tokens. The private preview hit 1 million. Then the 2 million figure went general. For comparison, many rival models in 2024 topped out at 128K to 200K. The jump was real. A legal team could feed a full discovery docket without chunking. A developer could ask for a review of a 10,000-file repository. It wasn't seamless; it was practical. And now it's history.
According to the same Gate.AI summary, rate limits for paid tiers once hit 1,000 requests per minute as of September 2024. That was enough for a mid-size team to run batch analysis on large video libraries. Today, that rate doesn't apply, because the endpoint is gone. The retirement isn't a warning. It's a done deal.
What 2 Million Tokens Actually Means for Businesses
A 2 million-token window isn't marketing language. It is a hard ceiling that changes project economics. For a Jakarta-based firm reviewing multi-year contracts in Indonesian and English, chunking thousands of pages into 8K or 32K blocks means losing cross-reference consistency. With 2 million tokens, the entire archive stays together. That saves engineering time, not just compute.
The catch is that long-context work isn't cheap. Historical pricing for Gemini 1.5 Pro included cached-context discounts, but the retired endpoint no longer carries a confirmed rate. That creates a planning headache. Companies that built workflows around the 2M window now have to migrate to Gemini 2.0 or 2.5 families, where pricing and token limits have shifted. It's not a simple upgrade. It's a rebuild.
Some teams are holding on to cached outputs or storing old results rather than re-running analysis. That works until the data changes. It doesn't scale. It's a temporary fix that hides a real transition problem.
A logistics company I spoke with last year ran fleet-sensor video uploads through Gemini 1.5 Pro to extract incident reports. Each video was two hours long, roughly 35,000 tokens of transcripts plus metadata. They queued 40 videos per day. At the old rate, that was manageable. After retirement, the same videos need an alternative model with a smaller window, which means splitting each one into segments. Accuracy dropped. They never said so publicly, but their engineering lead admitted the handoff logic now fails on edge cases that the long window used to absorb silently.
The Pricing Trap Nobody Talks About
There's a reason the pricing is vague. When a model is retired, providers stop publishing rates rather than maintaining obsolete charts. But businesses don't retire their dependencies at the same speed. A consultant in Surabaya who built a document-analysis pipeline on gemini-1.5-pro-002 in early 2025 is now facing a forced migration with no price anchor for comparison.
The real problem isn't the cost per token. It's the switching cost. New Gemini models offer different context windows, different multimodal rules, and different output formats. A pipeline that relied on the 2M window to hold an entire video transcript plus metadata now needs redesign, not just a new API key. That's where the silent tax lives.
Some teams are holding on to cached outputs or storing old results rather than re-running analysis. That works until the data changes. It doesn't scale. It's a temporary fix that hides a real transition problem.
Cloud billing teams hate this kind of migration. They can forecast spend on a published tiered table, but not on a moving target. Google's public docs for the 1.5 line stopped updating months before the official retirement. Internal Slack threads and partner briefings became the only reliable source of truth. That's a leak no enterprise buyer signs up for when they lock in a three-year commitment.
A fintech in Jakarta that I can't name told me it kept 1.5 Pro alive through a legacy project for six months after retirement, just to finish its regulatory reporting backlog. The finance team treated the extra spend as a write-off, not a budget line. It's the kind of decision that looks fine in a spreadsheet and terrible in a board deck.
Where Google Goes Next
Google hasn't replaced 1.5 Pro with a direct 2M-token clone. The newer families moved to different architectures and pricing structures, with context windows that vary by tier and model version. That makes sense technically, but it doesn't help the user who just wants the same long-context behavior at a known price.
Industry chatter, including from Google's own I/O notes, points to Gemini 2.5 Flash and Pro series as the real successors. Those models still support long contexts, but the public windows land in the low millions of tokens at best, and the pricing is consumption-based plus a separate caching tier. The shift favors big spenders who can pre-pay, not small teams that pay as they go.
Open-source alternatives are circling. The Llama 3.2 and 3.3 families pushed 400K-token windows last year, and Chinese vendors like MiniMax and StepFun quietly demoed multi-million-token runs at lower per-token rates. None match Google's tooling integration, but none need to. The market is fragmenting exactly when Google was consolidating it.
The lesson is practical. If you're running AI in Indonesia or anywhere else, treat long-context models as temporary infrastructure, not permanent fixtures. Document your prompt structures, save your chunking logic, and keep a migration budget. The 2 million-token window was a real advance, and the retirement is a reminder that quick progress doesn't mean stable contracts.
Buildings don't need foundation checks when the ground shifts, but codebases built on a single model endpoint do. Teams that assumed 2M tokens would stay cheap and available are now refactoring while traffic still flows. That's the actual cost of progress: not the API bill, but the re-architecture nobody planned for.
For more on AI industry pricing and regulation, see Reuters technology coverage. And for ongoing desk coverage of model releases and retirements, check our AI category page or read a related story on long-context models. The next shift won't wait for a price chart.