Models & Evaluation

Gemini 1.5 Pro

Gemini 1.5 Pro was a natively multimodal, sparse mixture-of-experts model from Google DeepMind, announced in February 2024 and notable for a context window of up to 1 million tokens (later 2 million). It anchored the Gemini family, which by 2026 had advanced through the 2.x and 3.x generations.

What Gemini 1.5 Pro Was

Gemini 1.5 Pro was a large multimodal language model announced by Google on February 15, 2024, as the first model in the Gemini 1.5 generation. It was a mid-sized member of the Gemini family, positioned for a wide range of tasks, and was distributed through Google AI Studio and Vertex AI for developers and enterprise customers, with consumer access arriving later through the Gemini app and Gemini Advanced subscription. Its headline feature was a very long context window: an initial preview allowed up to 1 million tokens, with a 2 million token window announced at Google I/O in May 2024 and opened to all developers on June 27, 2024.

The model was natively multimodal, meaning it was trained from the start to process text, images, audio, and video in a single interleaved input stream rather than bolting separate encoders onto a text-only core. The Google DeepMind technical report described it as a compute-efficient model that could recall and reason over fine-grained information across millions of tokens, including long documents, hours of video, and hours of audio. A 1 million token window could hold roughly one hour of video, eleven hours of audio, or more than thirty thousand lines of code.

Architecturally, Gemini 1.5 Pro was built on a sparse mixture-of-experts (MoE) Transformer. This let the total parameter count grow while keeping the number of parameters activated per token roughly constant, improving compute efficiency relative to a dense model of equivalent capacity. By June 2026 the model itself had been retired from the API (the gemini-1.5-pro alias was shut down on September 24, 2025), but it remains a historically important anchor for the long-context and multimodal capabilities that later Gemini generations inherited and extended.

Announced February 15, 2024 as the first Gemini 1.5 model.
Context window of up to 1 million tokens at launch, expanded to 2 million tokens by mid-2024.
Delivered via Google AI Studio and Vertex AI for developers, and the Gemini app for consumers.
Retired from the Gemini API in 2025; the gemini-1.5-pro alias shut down September 24, 2025.

The Gemini Tier Naming and How It Evolved

The Gemini brand launched in December 2023 with three size tiers: Gemini Ultra for highly complex tasks, Gemini Pro for general-purpose use, and Gemini Nano for on-device tasks running on hardware such as Pixel phones. The 1.0 generation framed these tiers primarily by capability and deployment target rather than by latency and cost.

Gemini 1.5, announced in February 2024, narrowed the public lineup over 2024 to Pro and a smaller, faster Flash variant (Flash arrived at Google I/O in May 2024). Flash was a lighter model distilled for lower latency and cost, while Pro carried the long-context and reasoning workload. This Pro-versus-Flash split became the durable spine of the family: Pro for capability, Flash for speed and price. The Ultra label was not carried forward as a routine API tier, and Nano continued to denote the on-device line separately.

Later generations refined the pattern. Gemini 2.0 arrived in December 2024 and Gemini 2.5 in March 2025, adding a Flash-Lite tier at the lowest cost and latency point. Gemini 3 was announced on November 18, 2025, and the family has since advanced to the Gemini 3.x generation: as of mid-2026 the current flagship is Gemini 3.1 Pro (released February 19, 2026), with Pro, Flash, and Flash-Lite as the standard tiers and additional specialized variants. Specific version numbers shift frequently, so the durable takeaway is the tier structure rather than any single point release.

Gemini 1.0 (December 2023) introduced Ultra, Pro, and Nano tiers.
The Gemini 1.5 generation (2024) standardized the Pro-for-capability, Flash-for-speed split.
Gemini 2.5 (2025) added a Flash-Lite tier for the lowest cost and latency.
Gemini 3 was announced November 18, 2025; by mid-2026 the current flagship was Gemini 3.1 Pro (February 19, 2026), with the 2.5 and 3.x lines as the recommended models.

Native Multimodality

Native multimodality means a model is trained jointly on multiple data types so that a single network processes text, images, audio, and video together, rather than routing each modality through an independent system. Gemini 1.5 Pro accepted interleaved inputs, so a single prompt could mix a transcript, a chart image, and a video clip, and the model could reason across all of them at once. This contrasts with pipeline approaches that transcribe audio or caption images with separate tools before a text-only model sees the result.

The long context window amplified this design. Because Gemini 1.5 Pro could hold up to a million or more tokens, entire films, lengthy audio recordings, or large codebases could be placed directly in the prompt. The technical report highlighted state-of-the-art results on long-document question answering, long-video question answering, and long-context automatic speech recognition, tasks that depend on holding large multimodal inputs in working context simultaneously.

Native multimodality also shaped how downstream applications were built. Instead of stitching together optical character recognition, speech-to-text, and a separate language model, developers could send raw media and rely on one model for understanding. This reduced integration overhead but also meant that errors in any modality propagated through the same model, so careful evaluation across modalities mattered.

Processed text, images, audio, and video in a single interleaved input stream.
Avoided separate captioning or transcription pipelines for each modality.
Reached strong results on long-document QA, long-video QA, and long-context speech recognition.
Long context let entire films, audio recordings, or codebases sit directly in the prompt.

Long Context and Its Caveats

The defining capability of Gemini 1.5 Pro was its long context window. Google DeepMind reported near-perfect retrieval, above 99 percent, on needle-in-a-haystack tests up to at least 1 million tokens, with the report describing continued strong recall when stress-tested toward 10 million tokens in research settings. This was a clear step beyond contemporaries such as Claude 2.1 at 200,000 tokens and GPT-4 Turbo at 128,000 tokens.

Strong synthetic retrieval scores do not guarantee strong reasoning over long inputs. A well-documented failure mode is lost-in-the-middle, where models attend reliably to information near the start and end of a long prompt but degrade on facts buried in the middle. Single-fact retrieval (find one needle) is also easier than multi-hop reasoning that must combine many scattered facts, and accuracy typically falls as the number of required facts and the input length both grow. Treating a large context window as a drop-in replacement for retrieval-augmented generation can therefore be a mistake.

There are also practical costs. The self-attention mechanism in a Transformer scales quadratically with sequence length, so doubling the context roughly quadruples the attention compute, which motivates efficiency work such as context caching to amortize repeated long prefixes. Caching itself has thresholds and minimums that can surprise developers who expect every prompt to benefit. The reliable pattern is to verify, on the actual task, that the model uses mid-context information rather than assuming a long window alone solves retrieval.

Attention(Q, K, V) = softmax(QKᵀ / √d_k) V,   cost ∝ O(n² · d)

Scaled dot-product self-attention. The QKᵀ term compares every token to every other token, so compute and memory grow as O(n²) in sequence length n (with model dimension d), the reason very long context is expensive and motivates caching and efficient-attention work.

Reported above 99 percent needle-in-a-haystack retrieval up to at least 1 million tokens.
Lost-in-the-middle: facts buried mid-prompt are recalled less reliably than those at the edges.
Single-fact retrieval is far easier than multi-hop reasoning over many scattered facts.
A long window is not a free substitute for retrieval-augmented generation on hard tasks.

Mixture-of-Experts Architecture

Gemini 1.5 Pro used a sparse mixture-of-experts (MoE) Transformer. In an MoE layer, the single feed-forward block of a dense Transformer is replaced by many parallel expert sub-networks plus a learned router. For each token, the router selects a small subset of experts (often the top-k) to process it, so only a fraction of the model's total parameters are active on any given token.

The practical benefit is that total parameter count, and therefore representational capacity, can scale up while the activated parameters per token, and therefore the per-token compute, stay roughly fixed. This decoupling of capacity from inference cost is why MoE designs recur across large frontier models. The router is typically trained with auxiliary load-balancing objectives so that experts receive comparable amounts of traffic and the model does not collapse onto a few experts.

MoE introduces its own engineering challenges, including routing instability, uneven expert utilization, and the memory cost of holding all experts even though only some run per token. Google DeepMind credited the MoE design, together with other architecture and data advances, for Gemini 1.5 Pro matching or exceeding the earlier Gemini 1.0 Ultra on many benchmarks while using less training compute.

An MoE layer routes each token to a small subset of expert sub-networks via a learned router.
Total parameters can grow while activated parameters per token stay roughly constant.
Auxiliary load-balancing losses keep expert traffic even and prevent router collapse.
DeepMind credited MoE for 1.5 Pro rivaling the larger 1.0 Ultra at lower training compute.

Calling the Gemini API in 2026

The recommended way to call Gemini from Python is the Google Gen AI SDK, installed as the google-genai package and imported as google.genai. It reached general availability in 2025 and replaced the older google-generativeai library for new development. The same SDK targets both the Gemini Developer API (with a key from Google AI Studio) and Vertex AI, which is the path most enterprises use for governance and quotas.

Because Gemini 1.5 Pro itself was retired from the API in 2025, current code should target a supported model such as a 2.5 or 3.x Pro or Flash variant. The interface is otherwise stable: create a client, call generate_content with a model identifier and the prompt contents, and read the text response. Multimodal inputs are passed as additional content parts (uploaded files or inline media) alongside text.

When working with very long inputs, the API supports context caching so that a large shared prefix is processed once and reused across requests, reducing cost and latency. Caching has minimum token thresholds and time-to-live settings, so it pays to confirm a given prompt is actually large enough to benefit before relying on it.

python

from google import genai

# Reads GEMINI_API_KEY from the environment (key from Google AI Studio)
client = genai.Client()

response = client.models.generate_content(
    model="gemini-2.5-flash",  # 1.5 Pro is retired; use a current model
    contents="Summarize the key idea of mixture-of-experts in one sentence.",
)

print(response.text)

A minimal call with the current Google Gen AI SDK (google-genai). The same client.models.generate_content interface accepts multimodal contents and any supported model identifier.

Use the google-genai SDK (import google.genai); it superseded google-generativeai.
One SDK targets both Google AI Studio keys and Vertex AI deployments.
Target a current supported model; the gemini-1.5-pro identifier no longer resolves.
Context caching reuses a long shared prefix but has token-count minimums and TTLs.

Key takeaways

Gemini 1.5 Pro, announced February 15, 2024, was a natively multimodal sparse mixture-of-experts model from Google DeepMind.
Its context window reached up to 1 million tokens at launch and 2 million tokens by mid-2024, a large jump over contemporaries.
Near-perfect needle-in-a-haystack retrieval did not eliminate lost-in-the-middle effects or multi-hop reasoning failures at long context.
Self-attention scales as O(n²) in sequence length, which makes long context expensive and motivates context caching and efficient-attention work.
The Pro-for-capability and Flash-for-speed tier split, introduced with 1.5, persisted through the 2.x and 3.x generations released by 2026.
The model was retired from the API in 2025; current work should target a supported 2.5 or 3.x model via the google-genai SDK.

Frequently asked questions

Gemini 1.5 Pro launched in February 2024 with a context window of up to 1 million tokens in preview. At Google I/O in May 2024 a 2 million token window was announced, and it was opened to all developers on June 27, 2024. A 1 million token window holds roughly an hour of video, eleven hours of audio, or more than thirty thousand lines of code.

No. The Gemini 1.5 models were deprecated and the gemini-1.5-pro alias was shut down on September 24, 2025. By June 2026 the recommended models on the Gemini API span the 2.5 and 3.x generations. New projects should target a current Pro, Flash, or Flash-Lite model instead.

Gemini 1.5 Pro is a sparse mixture-of-experts (MoE) Transformer. An MoE layer routes each token to a small subset of expert sub-networks, so the model's total parameters can grow while the parameters activated per token stay roughly constant. Google DeepMind credited this design for matching the larger Gemini 1.0 Ultra at lower training compute.

Not reliably. Although Gemini 1.5 Pro reported above 99 percent recall on single-needle retrieval tests, long-context models still show lost-in-the-middle degradation and struggle with multi-hop reasoning over many scattered facts. For hard tasks it is safer to verify mid-context performance rather than assume a large window alone solves retrieval.

Use the Google Gen AI SDK, installed as the google-genai package and imported as google.genai. It reached general availability in 2025 and replaced the older google-generativeai library. The same SDK works with both Google AI Studio API keys and Vertex AI deployments.

Pro models target the highest capability and reasoning, while Flash models are smaller and faster, optimized for lower latency and cost. The 1.5 generation established this split in 2024, and later generations added a Flash-Lite tier at the lowest cost and latency point. The right choice depends on whether a task needs maximum quality or maximum throughput.

Put the idea into practice

MemX is an AI memory agent built on these ideas: store anything, skip the folders, and find it again by asking in plain English.

Try MemX Free