Gemini 2.5 Pro

Gemini 2.5 Pro is the flagship reasoning large language model of Google's Gemini 2.5 family, developed by Google DeepMind and first released as an experimental preview on March 25, 2025, under the model identifier gemini-2.5-pro-exp-03-25. ^[1] Google described it at launch as "our most intelligent AI model," the first generation in which all Gemini models are "thinking models, capable of reasoning through their thoughts before responding," meaning every response begins with an internal chain of reasoning before any text is shown to the user. ^[1] Gemini 2.5 Pro debuted at number one on the LMArena (Chatbot Arena) human-preference leaderboard, with a debut score of 1,443, close to 40 points ahead of Grok 3 and GPT-4.5 ^{[33] [34]}, and reached general availability on June 17, 2025, priced at $1.25 per million input tokens and $10.00 per million output tokens at the standard tier. ^{[1] [3] [5]} It is offered through Google AI Studio, the Gemini API, and Vertex AI. ^[3]

At launch the model posted state-of-the-art scores on Humanity's Last Exam (18.8 percent without tools), GPQA Diamond (84.0 percent), and AIME 2025 (86.7 percent), and pulled level with or ahead of Anthropic's Claude 3.7 Sonnet and OpenAI's GPT-4.5 across most public reasoning benchmarks. ^{[1] [15]} It also took the top position on WebDev Arena for full web app generation. The model's broad capability profile, 1 million token context window, and standard pricing of $1.25 per million input tokens and $10.00 per million output tokens led to fast adoption inside developer tools such as Cursor, Replit, and Cognition's Devin through the spring and summer of 2025.

The Gemini 2.5 family also includes the smaller Gemini 2.5 Flash and Gemini 2.5 Flash-Lite siblings, plus an enhanced reasoning configuration called Gemini 2.5 Pro Deep Think, announced at Google I/O on May 20, 2025. ^[2] Deep Think went on to achieve a gold-medal score at the 2025 International Mathematical Olympiad in July and at the International Collegiate Programming Contest World Finals in September, both under official competition rules. ^{[26] [27]} Gemini 2.5 Pro was eventually superseded by Gemini 3 Pro in November 2025, but it remained the default "Pro" model on the Gemini API, in the Gemini consumer app's paid tiers, and across Google products for nearly eight months and continues to be served as a price-performance option into 2026. ^[4]

Background

Gemini 2.5 Pro is the third generation of Google's flagship Gemini line. It follows Gemini 1.5 Pro (February 2024) and the short-lived Gemini 2.0 Pro Experimental (February 2025), and it is the first Gemini model in which reasoning is the default rather than a separate variant.

Gemini 1.5 Pro arrived in February 2024 with a 1 million token context window, later expanded to 2 million tokens for waitlisted developers. It was the first Gemini model to use a Mixture-of-Experts (MoE) architecture, where only a subset of expert sub-networks is activated for each token, and it supported native input across text, images, audio, video, and code. Gemini 1.5 Pro introduced the very long context window that would come to define the Gemini line, but it had no dedicated chain-of-thought mode.

Gemini 2.0 Flash Experimental was announced on December 11, 2024, and reframed Google's strategy around what the company called the "agentic era." It added native tool use (Google Search, code execution), native audio output, real-time streaming through the Multimodal Live API, and roughly twice the speed of 1.5 Pro at lower cost. Eight days later, on December 19, 2024, Google released the experimental Gemini 2.0 Flash Thinking, a variant trained to spend additional inference compute on internal deliberation before answering. Flash Thinking foreshadowed the architectural direction of the 2.5 family.

In February 2025, Google released a wave of Gemini 2.0 stable models: an updated Gemini 2.0 Flash in the Gemini app in late January 2025, then on February 5 general API availability of 2.0 Flash alongside Gemini 2.0 Pro Experimental and a public preview of Gemini 2.0 Flash-Lite. ^[11] Gemini 2.0 Pro Experimental shipped with a 2 million token context window, native tool use, and stronger coding performance, but no built-in reasoning step. It was a brief stop on the way to 2.5.

The broader market context for the 2.5 launch was unusually crowded. DeepSeek R1, an open-weight reasoning model from a Chinese research lab, had landed in late January 2025 and demonstrated that distilled chain-of-thought training could produce frontier-class results at a fraction of the cost. ^[23] xAI released Grok 3 on February 17, 2025, and announced its own Big Brain reasoning mode. ^[22] OpenAI released GPT-4.5 on February 27, 2025, as a non-reasoning frontier model with strong general knowledge but uneven math and code performance. ^[21] Anthropic released Claude 3.7 Sonnet on February 24, 2025, which became the first commercial model to expose a configurable extended thinking mode to all developers. ^[20] By the time Gemini 2.5 Pro arrived four weeks later, every major lab had committed to some form of test-time reasoning. Google's contribution was to make thinking the default rather than a toggle.

When was Gemini 2.5 Pro released?

Gemini 2.5 Pro was announced on March 25, 2025, in a Google DeepMind blog post titled "Gemini 2.5: Our newest Gemini model with thinking," written by Google DeepMind CTO Koray Kavukcuoglu. ^[1] The post described the model as "our most intelligent AI model" and framed the 2.5 series as a generation in which all Gemini models will be thinking models, capable of reasoning through their own thoughts before producing a response. ^[1]

The initial release shipped under the experimental model ID gemini-2.5-pro-exp-03-25. It was made available immediately and at no cost in Google AI Studio, and rolled out the same day to the Gemini consumer app for Gemini Advanced subscribers. ^[1] Google said Vertex AI access would follow in the coming weeks. ^[1] The free tier included generous rate limits intended to encourage broad early experimentation, and developer documentation listed a 1 million token context window with an announced ("coming soon") expansion to 2 million tokens. ^[1]

Google's launch claims fell into three categories. First, the model debuted at number one on LMArena (the human-preference leaderboard run by the LMSYS / Chatbot Arena project), with a reported lead of close to 40 Elo points over the second-place model. ^{[1] [18]} LMArena listed the launch score at 1,443, and Arena described the margin as the largest jump it had recorded. ^{[33] [34]} Second, Google reported industry-leading scores on Humanity's Last Exam (18.8 percent without tools), GPQA Diamond (84.0 percent), and AIME 2025 (86.7 percent), along with strong but not market-leading results on SWE-bench Verified (63.8 percent). ^[1] Third, the company emphasized improvements in coding: a top placement on the WebDev Arena leaderboard for full web app generation, and a 70.4 percent score on LiveCodeBench v5 for competitive programming.

The "Pro Experimental" label was important. Google had used the phrase "Pro Experimental" earlier in 2025 for Gemini 2.0 Pro Experimental, signaling that a model was production-eligible in spirit but still subject to changes in pricing, rate limits, and behavior before stable release. Gemini 2.5 Pro Experimental sat in this category until April 4, 2025, when a paid preview (gemini-2.5-pro-preview-03-25) opened in the Gemini API. ^[31] The model received an updated checkpoint, gemini-2.5-pro-preview-05-06, just before Google I/O 2025 (May 20-21), positioned specifically as a coding upgrade. ^[6] The 05-06 update also reduced errors in function calling. ^[6] A further refresh, gemini-2.5-pro-preview-06-05, raised the model's LMArena score by 24 points to 1,470 and its WebDev Arena score by 35 points to 1,443, and improved style and formatting. ^[35] The June 17, 2025, GA release retired the preview suffix in favor of the stable identifier gemini-2.5-pro. ^[3]

Variants and snapshot timeline

Google shipped multiple variants and snapshots of Gemini 2.5 in 2025. The table below summarizes the major identifiers as they appeared in the Gemini API and Vertex AI catalogs. ^[9]

Gemini 2.5 Pro

The headline model. It uses the largest parameter budget of the 2.5 family and has the longest configurable thinking budget (up to 32,000 tokens of internal reasoning). ^[7] It is the only Gemini 2.5 model approved for the Deep Think configuration and the only one with the upgraded coding RL training that drove the WebDev Arena number-one finish.

Gemini 2.5 Flash

First Flash-class model with built-in thinking. It shares the same 1 million token context window as Pro and supports the same input modalities, but uses a smaller and faster expert routing configuration. Google reported that the I/O 2025 preview of Flash used 20 to 30 percent fewer tokens than its predecessor on internal evaluations while improving on reasoning and code benchmarks. ^[2] Flash reached general availability on June 17, 2025.

Gemini 2.5 Flash-Lite

The smallest and cheapest model in the 2.5 family. Flash-Lite is targeted at high-volume, latency-sensitive workloads such as text classification, translation, content moderation, and intelligent routing. Google described it as roughly 1.5 times faster than Gemini 2.0 Flash at lower cost. Flash-Lite entered preview on June 17, 2025, and reached general availability on July 22, 2025. ^[3]

Gemini 2.5 Pro Deep Think

Deep Think is an enhanced reasoning configuration of Gemini 2.5 Pro that uses additional inference techniques to consider multiple hypotheses in parallel before committing to a final answer. ^[2] It was announced at Google I/O on May 20, 2025, and described by Google as achieving an impressive score on the 2025 USA Mathematical Olympiad (USAMO). ^[2] Public reports cited a 49.4 percent score on USAMO 2025 and number-one placement on LiveCodeBench's competition-coding leaderboard.

Deep Think was held back for an extended safety review and rolled out first to trusted testers via the Gemini API; broader access for Google AI Ultra subscribers in the Gemini app began on August 1, 2025. ^[25] The consumer release was a faster, more usable variant of the same line of work that won the IMO gold medal: Google reported it reached bronze-level performance on the 2025 IMO benchmark and scored 87.6 percent on LiveCodeBench competition-level coding, up from 80.4 percent in May. ^[25] In July 2025, a research version of Deep Think achieved an official gold-medal score at the 2025 International Mathematical Olympiad, solving five of six problems for 35 out of 42 possible points, working entirely in natural language and within the 4.5-hour competition limit per session. ^[26] In September 2025 the same line of work reached gold-medal level at the International Collegiate Programming Contest World Finals, solving ten problems including one (Problem C) that no human team solved during the contest. ^[27]

Gemini 2.5 Pro TTS Preview

A speech-synthesis-focused variant, gemini-2.5-pro-preview-tts, released in preview on May 20, 2025. It is served through the standard Gemini API; its model page lists the Live API as not supported. ^[30][31] It supports controllable speaker voices, emotional inflection, and over 24 languages. The TTS variant is optimized for structured audio output (audiobook, podcast, narration) rather than open-ended chat. As of October 2026, Google's deprecations page lists a shutdown date of November 17, 2026, for it and names Gemini 3.8 Flash TTS or Gemini 3.8 Flash-Lite TTS as the recommended replacement. ^[29]

Architecture

Google DeepMind has not published a full technical report disclosing the parameter count or training corpus of Gemini 2.5 Pro. The technical brief released in mid-2025 ("Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context") describes the model qualitatively rather than in terms of architecture details. ^[7] From the brief and from independent commentary, the following design choices are well established.

Gemini 2.5 Pro is a decoder-only Transformer built on the same architectural family as Gemini 1.5 and Gemini 2.0. It uses a Mixture-of-Experts routing scheme inherited from Gemini 1.5, with sparsely activated expert sub-networks chosen per token. The total parameter count has not been disclosed, and Google has not disclosed the active or total parameter count. ^[16]

The model is natively multimodal in the same sense as earlier Gemini releases: text, image, audio, video, and PDF inputs are converted to a shared token representation by per-modality encoders, then processed by a single Transformer stack. There is no separate vision tower running on the side. This is in contrast to systems that bolt a vision encoder onto a text-only model after the fact.

The defining 2.5 change is the integration of reasoning into the default response path. Rather than a separate "thinking" model, every Gemini 2.5 Pro request runs through an internal chain-of-thought pass. Developers can set a thinking budget (in tokens) that caps how much computation the model spends on internal deliberation. Unlike Gemini 2.5 Flash, Gemini 2.5 Pro does not allow thinking to be turned off: its budget ranges from 128 to 32,768 tokens. ^[37] The default thinking budget is dynamic, based on Google's internal heuristics. Reasoning is trained through a combination of supervised fine-tuning on chain-of-thought traces and reinforcement learning from preference data, drawing on techniques explored earlier in Gemini 2.0 Flash Thinking and on research published by Google DeepMind throughout 2024.

The model supports inspectable thought summaries: a redacted form of its internal reasoning that developers can return alongside the final answer. Google added the feature to the API around Google I/O 2025, presenting thoughts with headers, key details, and explicit tool-call annotations so developers can audit how the model arrived at an answer. ^[19] Full chain-of-thought transcripts are not returned; developers receive the summaries instead.

The knowledge cutoff is January 2025 for the GA model, as listed in Google's technical report. ^[7]

Capabilities

Reasoning

Gemini 2.5 Pro's reasoning behavior is most visible in long, multi-step problems. Internal Google evaluations described in the launch blog show the model decomposing graduate-level science problems, working through multi-step proofs in mathematics, and reading lengthy code diffs before proposing a fix. ^[1] Developers can ask the model to summarize its thinking, which yields a paragraph or two of human-readable reasoning that can be used for debugging or transparency.

The configurable thinking budget (up to 32,000 tokens) lets applications trade latency for accuracy. Internal Google plots show smooth improvements on AIME 2025, LiveCodeBench, and GPQA Diamond as the budget increases. ^[7] The diminishing-returns point depends on the task; for many real-world coding queries, a budget of a few thousand tokens captures most of the benefit.

Coding and software engineering

Google invested heavily in coding for the 2.5 series. The May 6, 2025, preview snapshot (gemini-2.5-pro-preview-05-06) was specifically positioned as a coding upgrade, and Google's developer blog reported large improvements on the company's internal evals. ^[6] Public results that landed in the Vellum, Helicone, and Artificial Analysis comparisons placed Gemini 2.5 Pro at or near the top on most coding benchmarks: WebDev Arena (number one for full app generation), Aider Polyglot (74.0 percent on real-world code editing across multiple languages), and LiveCodeBench v5 (70.4 percent on competitive programming). ^{[15] [16]}

On SWE-bench Verified, the standard benchmark for autonomous bug-fixing on real GitHub repositories, the model scored 63.8 percent at launch using a custom agent harness. ^[1] That number trailed Claude 3.7 Sonnet's 70.3 percent (with Anthropic's custom scaffold) on the same benchmark, and was acknowledged as such by Google. SWE-bench Verified remained the one major coding benchmark where Anthropic's Claude line held a clear lead through mid-2025.

Qualitatively, Replit, Cursor, and Cognition all worked with Gemini 2.5 Pro within weeks of the preview. ^[6] Replit president Michele Catasta said his team found it "the best frontier model when it comes to 'capability over latency' ratio," and Silas Alberti of Cognition said the updated model "felt like a more senior developer." ^[6] Cursor users widely adopted the model for refactoring and multi-file edits, particularly in front-end work where the WebDev Arena strength translated to better UI generation.

Multimodal understanding

Gemini 2.5 Pro accepts text, images, audio, video, and PDFs in the same prompt. ^[4] Concrete capacity limits include up to 3,000 images per prompt (each up to 7 MB inline or 30 MB through Cloud Storage), approximately 45 minutes of video with audio (or one hour without), up to roughly 8.4 hours of audio per prompt, and PDFs of up to 3,000 pages or 50 MB per file on Vertex AI. ^[4]

Video understanding is one of the model's most differentiated capabilities. With the May 2025 preview (I/O edition), Google reported a score of 84.8 percent on VideoMME, a standard benchmark for multimodal video comprehension, and showed a "Video to Learning App" demo in which the model turned a single YouTube video into an interactive learning app. ^[6] The capability was widely shared on social media and was one of the reasons Gemini 2.5 Pro became the preferred model for creative and educational video tasks in mid-2025.

Audio handling is also strong. The model can transcribe speech, describe ambient sounds, identify musical instruments, and reason about spoken content over hours of input. The Live API exposes a real-time streaming mode that supports interruptions and turn-taking. A dedicated TTS preview offers controllable voice synthesis across more than 24 languages.

Long context

Gemini 2.5 Pro launched with a 1 million token context window. Google announced at the same time that 2 million tokens were "coming soon," matching the eventual ceiling of Gemini 1.5 Pro. ^[1] The 2 million token expansion did not roll out to general availability for Gemini 2.5 Pro through 2025 or into early 2026; the ceiling remained at 1 million tokens at GA, then through the model's life as the default Gemini Pro through November 2025 and beyond. ^[4]

A 1 million token window is enough for roughly 750,000 words of English text, several hours of audio, or a multi-thousand-line codebase. On the MRCR (Multi-Round Coreference Resolution) benchmark at 128k context, the model scored 94.5 percent, indicating strong long-context retrieval. ^[36] Citizen Health, an enterprise customer, said it streams decades of longitudinal patient records, including clinician notes, imaging reports, and genomic panels, into the model's million-token context window. ^[3]

Agentic and tool use

Gemini 2.5 Pro supports function calling, structured outputs, batch processing, and context caching, all of which are important for building AI agents that interact with external tools and APIs over multiple steps. Google extended the API at I/O 2025 with thought summaries, asynchronous Batch endpoints at roughly 50 percent of interactive cost, and broader native tool use including code execution and Google Search grounding. ^[19] The model also supports the Model Context Protocol (MCP) standard for connecting to data sources, and Google integrated Project Mariner's computer-use capabilities into the Gemini 2.5 API as a preview, which allowed the model to control a web browser. ^[19]

Benchmarks

The table below summarizes the public benchmark numbers Google reported at and shortly after the March 2025 launch, along with the competitor figures Google listed in the same comparison (mostly self-reported by each provider; multiple-attempt scores are marked). ^[36]

A few comparisons are worth noting. Claude 3.7 Sonnet held a clear lead on SWE-bench Verified, the headline benchmark for agentic coding. Grok 3's extended-thinking mode beat Gemini 2.5 Pro on AIME 2025 only with multiple attempts (93.3 percent; 77.3 percent single-attempt), and trailed it on broader reasoning benchmarks. GPT-4.5 was strong on world-knowledge benchmarks like SimpleQA but weak on competition mathematics, since it was a non-reasoning model. DeepSeek R1, despite being open-weight and substantially cheaper to serve, was not competitive on multimodal or long-context tasks. On ARC-AGI-2, the abstract-reasoning grid benchmark introduced in 2025, every model in this comparison scored in the low single digits to mid-single digits; Gemini 2.5 Pro's 4.9 percent was typical of the frontier cohort and underscored that the benchmark remained largely unsolved before Gemini 3.

Gemini 2.5 Pro Deep Think pushed several of these numbers further. Public reporting cited a 49.4 percent score on USAMO 2025, an 84.0 percent score on MMMU at the Deep Think setting, and a top-of-leaderboard finish on LiveCodeBench's competition coding category. ^[2] Deep Think also led the FrontierMath benchmark in its tier 1 to 3 range with approximately 29 percent. The most-cited Deep Think result came in July 2025, when an advanced research version solved five of six problems for 35 out of 42 points at the 2025 International Mathematical Olympiad, working in natural language within the standard 4.5-hour competition limit and earning an official gold-medal score. ^[26] IMO President Prof. Dr. Gregor Dolinar confirmed the grade, stating that "Google DeepMind has reached the much-desired milestone, earning 35 out of a possible 42 points, a gold medal score," and that the model's solutions were "clear, precise and most of them easy to follow." ^[26] Two months later, Deep Think reached gold-medal level at the ICPC 2025 World Finals, solving ten problems including one that no human team solved; by combined solve time it would have ranked second overall against the university teams. ^[27]

Independent evaluations through 2025 reinforced the launch numbers but added important nuance. METR's time-horizon measurements put the June 2025 preview of Gemini 2.5 Pro at a 50 percent time horizon of about 40 minutes, behind Claude 3.7 Sonnet (about 55 minutes) and OpenAI's o3 (about 91 minutes). ^[17] Simon Willison, testing the model on launch day, found its coding results comparable to his previous favourite, Claude 3.7 Sonnet. ^[13] By early 2026, Artificial Analysis tracked the released gemini-2.5-pro model at an Intelligence Index score of 35 and a blended price of $3.44 per million tokens, ranking it as a mid-tier reasoning model relative to the Gemini 3 and GPT-5 cohort that followed. ^[16]

How much does Gemini 2.5 Pro cost?

Gemini 2.5 Pro is billed per million tokens through both the Gemini API (Google AI Studio) and Vertex AI. ^[5] Pricing is tiered by prompt size: requests with up to 200,000 input tokens are charged at the standard rate, while requests with longer prompts use a higher per-token rate. Multiple service tiers exist. The free rate-limited tier in Google AI Studio is intended for experimentation and small projects, and paid Standard, Batch (also called Flex), and Priority tiers cover production workloads.

Gemini 2.5 Pro pricing (per million tokens)

Context caching is available at $0.125 per million tokens for prompts under 200k tokens ($0.25 above 200k), with cache storage at $4.50 per million tokens per hour. ^[5]

Gemini 2.5 Flash pricing (per million tokens)

Gemini 2.5 Flash-Lite pricing (per million tokens)

Gemini 2.5 Pro's $1.25 input and $10.00 output Standard pricing put it well below GPT-4.5 (which OpenAI priced at $75 per million input tokens and $150 per million output tokens, more than ten times higher) and roughly in line with Claude 3.7 Sonnet ($3 / $15 per million tokens). Pricing for the long-context tier (above 200k input tokens) doubled the input cost and increased output cost by 50 percent, similar to how Anthropic and OpenAI have approached very long contexts. The Batch / Flex tier offered a roughly 50 percent discount for asynchronous workloads where latency was not critical.

Availability extended across three primary surfaces. Free experimentation lived in Google AI Studio, with rate limits of a few requests per minute on the experimental endpoint. Production use ran through the paid Gemini API or through Vertex AI for enterprise customers needing fine-tuning, audit logs, and Google Cloud integrations. ^[3] Consumer access came through the Gemini app's paid tiers (Gemini Advanced, then later Google AI Pro at $19.99 per month and Google AI Ultra at $249.99 per month after Google's May 2025 subscription restructuring); Deep Think was gated to the Ultra tier when it rolled out to the consumer app in August. ^[25]

Adoption

Third-party adoption began within hours of the March 25 launch. Cursor added the experimental endpoint as a selectable model the same week and benchmarked it as one of the top performers on its internal coding evals. Replit integrated the model into Replit Agent and Ghostwriter; the company publicly described the integration as their preferred frontier model for agent workflows. Aider, the open-source command-line code editor, included it in the polyglot leaderboard, where it took the top spot for several weeks.

In the broader developer ecosystem, the model became a default choice for chat applications and tools that wanted reasoning at low cost. LangChain, LlamaIndex, and the Vercel AI SDK shipped first-class support within the first month. Companies building AI agents layered Gemini 2.5 Pro into pipelines that previously required calls to OpenAI o1 or Claude 3.7 Sonnet, citing the lower price and the larger context window.

Enterprise adoption ran along two tracks. Vertex AI customers used it for code generation, document understanding, and translation across long enterprise documents. Citizen Health used it on patient records. Cognition (the company behind Devin) integrated it as one of several backend models, comparing its performance against Anthropic and OpenAI counterparts on real customer workloads. By early 2026 Google reported more than 8 million Gemini Enterprise seats sold across more than 2,800 companies, and much of that 2025 momentum traced back to Gemini 2.5 Pro as the default high-end model during the year. ^[28]

Inside Google, Gemini 2.5 Pro became the default "Pro" model in the Gemini app's paid tiers, took over reasoning queries in Search's AI Mode, and provided the underlying intelligence for Project Astra (Google's universal-assistant prototype) and Project Mariner (its browsing agent). ^[2] Workspace customers saw Gemini 2.5 Pro power document-aware chat in Gmail, Docs, and Drive, particularly for summarization tasks that benefit from the long context window.

Reception

The initial March 25 release drew substantial attention from the developer community. Its 40-point lead on LMArena, which Arena called the largest jump it had seen, was widely noted. ^[34] Simon Willison, in his March 25 write-up, called it "a very strong new model" and said its coding results felt comparable to his previous favourite, Claude 3.7 Sonnet. ^[13]

The Vellum AI coding leaderboard, run by the AI development platform of the same name, had Gemini 2.5 Pro and Claude 3.7 Sonnet trading the top spot through April.

METR's task suite measures the time horizon over which an agent can autonomously complete real software tasks. Its measurement of the June 2025 preview of Gemini 2.5 Pro, about 40 minutes at 50 percent reliability, put the model behind Claude 3.7 Sonnet and OpenAI's o3 but ahead of DeepSeek R1 (about 28 minutes). ^[17]

Critics pointed to occasional refusals on benign prompts containing sensitive keywords, hallucinated quotes from documents in long-context queries, and a regression in response quality observed on certain checkpoints in mid-2025. A Google AI Developers Forum thread, "Gemini 2.5 Pro's Response Quality Regression," opened in June 2025, collected user reports that the stable release hallucinated more than the 05-06 preview. ^[12]

Reception inside academia was generally positive. The model was adopted as a baseline in many subsequent reasoning-model papers, often replacing GPT-4o or Claude 3 Opus as the stand-in for a frontier closed model. Researchers cited the publicly available thought summaries as a useful diagnostic tool for understanding chain-of-thought failures, even though the redacted form prevents direct study of the underlying reasoning trace.

Deep Think's July 2025 IMO gold medal and September 2025 ICPC gold medal generated unusually broad press coverage for an AI model achievement, in part because both results were obtained under the actual competition's rules rather than a research-only setting. ^{[26] [27]} Google noted that, unlike its 2024 AlphaProof system, the model worked end-to-end in natural language without translation into a formal language such as Lean. ^[26]

How does Gemini 2.5 Pro compare with Claude 3.7 Sonnet and GPT-4.5?

The table below summarizes the major contemporary frontier models that Gemini 2.5 Pro competed with through early 2025.

Gemini 2.5 Pro had a distinctive competitive position. It was the only model in this group to combine top-tier reasoning, native video understanding, and a 1 million token context window in a single endpoint. Its weakness against Claude 3.7 Sonnet was concentrated on SWE-bench Verified and on certain long-horizon agent tasks. Its weakness against GPT-4.5 was concentrated on factual and conversational tasks where extra reasoning sometimes hurt rather than helped .

Gemini 2.5 Pro's pricing and free-tier access made it the most accessible frontier reasoning model for hobbyists and small teams in 2025. Through Google AI Studio, a developer could use the model at no cost within rate limits, with no credit card. That was a meaningful difference compared to Claude 3.7 Sonnet (which required an Anthropic account and paid credits for any sustained usage) and to GPT-4.5 (which was paid only and the most expensive model on the market by per-token pricing).

The next-generation Claude 4 models from Anthropic, Claude Opus 4 and Claude Sonnet 4, released in May 2025, raised the bar on agentic coding (Claude Opus 4 reached 72.5 percent on SWE-bench Verified at launch, and Claude Sonnet 4 reached 72.7 percent). They overtook Gemini 2.5 Pro on most coding benchmarks but did not match its multimodal breadth or long-context ceiling. Through the second half of 2025, the practical choice between Gemini 2.5 Pro and the Claude 4 line came down to workload: video and long-context tasks favored Gemini, agentic coding favored Claude.

Safety and limitations

Google DeepMind described the safety work for Gemini 2.5 Pro under the rubric of its Frontier Safety Framework, the company's internal policy for tracking and mitigating risks from advanced AI systems. ^[25] The framework defines critical capability levels for specific risk categories (such as cyber-offense, autonomy, and biological weapon uplift) and triggers internal mitigations when a model approaches one of these thresholds. For Gemini 2.5 Pro, Google's technical report said the model was evaluated for CBRN, cybersecurity, machine learning R&D, and deceptive-alignment capabilities and did not reach any Critical Capability Level. ^[7]

The Deep Think variant was held back specifically for additional safety review. Google's I/O announcement noted that Deep Think would be released to trusted testers first while the company conducted further frontier-safety evaluations, an unusually conservative posture for a Google model release that drew positive comment from the AI safety community. Public access to Deep Think did not roll out broadly until August 1, 2025, when it became available to Google AI Ultra subscribers in the Gemini app.

Known limitations carried by Gemini 2.5 Pro through its life cycle included the following.

The knowledge cutoff is January 2025. Events, publications, and data after that date are not in the model's weights. Tool use (Google Search, code execution) is the recommended workaround for queries that need post-cutoff information. ^[4]

Very long prompts produce noticeable time-to-first-token delays. Prompts in the high hundreds of thousands of tokens can take tens of seconds to begin streaming, which makes them unsuitable for interactive applications. Artificial Analysis measured a typical time-to-first-token of roughly 27 seconds in late 2025, noting that this was on the high end for reasoning models in the same price tier. ^[16]

The announced 2 million token context expansion did not reach general availability for Gemini 2.5 Pro. The 1 million token ceiling was the practical maximum throughout the model's life as the default Pro endpoint, and remained so into 2026. ^[4]

Like all large language models, Gemini 2.5 Pro can hallucinate. The most-discussed real-world failure modes in mid-2025 included fabricated quotes from documents passed in long context, mixed-up details when reasoning across multiple PDFs in a single prompt, and confident citations of sources that were not present in the input. Users reported such regressions after the June 2025 stable release. ^[12]

Safety filters occasionally declined benign prompts that contained sensitive keywords. The defaults are configurable through the API, but the filter set is enforced more strictly in consumer-facing surfaces (the Gemini app, AI Overviews) than in developer-facing surfaces.

Deep Think adds latency and token cost. It is not suited to latency-sensitive production applications, and some Deep Think evaluations require explicit opt-in beyond the standard gemini-2.5-pro-deep-think model identifier.

The model also scored only 4.9 percent on ARC-AGI-2, the abstract-reasoning grid benchmark introduced in 2025 by Francois Chollet and the ARC Prize Foundation, highlighting that performance on conventional academic benchmarks did not translate to the kind of novelty-driven puzzle reasoning ARC-AGI-2 was designed to measure.

Finally, the model's output is text-only by default. Image, audio, and video generation are handled by separate Google models such as Imagen 3, Veo 2, and the Live API's native audio mode. Native multimodal output, present in Gemini 2.0 Flash for some modalities, was not part of the 2.5 Pro release.

Successor

Gemini 3 Pro launched on November 18, 2025, replacing Gemini 2.5 Pro as Google's flagship. The 3 Pro release came with a substantial leap on most benchmarks: 1,501 LMArena Elo, 91.9 percent on GPQA Diamond, 76.2 percent on SWE-bench Verified, and 37.5 percent on Humanity's Last Exam without tools. It kept the 1 million token context window and the thinking-by-default architecture, and added persistent memory and stronger agentic behavior. Pricing also rose, with Gemini 3 Pro charging $2 per million input tokens and $12 per million output tokens at standard context, compared to $1.25 / $10.00 for Gemini 2.5 Pro.

Gemini 2.5 Pro remained available on the Gemini API and Vertex AI after the Gemini 3 launch, both as a price-performance option for production workloads that did not need the extra capability and as a fallback during the rollout of Gemini 3. ^[29] As of October 2026, Google's Gemini API deprecations page lists no shutdown date for gemini-2.5-pro or gemini-2.5-flash, but limits the 2.5 models to developers who have used them before and points new projects to newer models such as Gemini 3.8 Flash. ^[29] On Vertex AI, Google lists October 20, 2026 as the retirement date for support of gemini-2.5-pro. ^[4]

The Flash and Flash-Lite siblings followed parallel transitions. Gemini 3 Flash launched in December 2025 and became the new default in the Gemini app, while Gemini 3.1 Pro Preview and Gemini 3.1 Flash Lite arrived in February and March 2026, respectively. The 2.5 line as a whole stepped into a long-tail role, supporting cost-sensitive and latency-sensitive workloads while the 3 series took over the frontier.

Continue reading

• Gemini (language model)Gemini is a family of natively multimodal large language models developed by Google DeepMind, first announced on December 6, 2023, that can reason across text, images, audio, video, and code within a single…Mentioned 161 times above
• Claude 3.7 SonnetClaude 3.7 Sonnet is a large language model developed by Anthropic and released on February 24, 2025Mentioned 16 times above
• Gemini 3Gemini 3 is the third major generation of the Gemini family of multimodal models from Google DeepMind, launched on November 18, 2025 with Gemini 3 Pro as the flagship and described by Google as "our most…Mentioned 13 times above
• SSWE-benchSWE-bench is an execution-based benchmark for evaluating whether a language-model system can resolve real software issues.Mentioned 9 times above
• Google DeepMindGoogle DeepMind is an artificial intelligence research and model-development organization at Google.Mentioned 8 times above
• Gemini 2.5 FlashGemini 2.5 Flash is a fast, cost-optimized multimodal large language model developed by Google DeepMind and the mid-tier member of the Gemini 2.5 family.Mentioned 7 times above

More from See also

• Google Vertex AI
• Google AI Studio
• Multimodal Model
• Mixture of Experts (MoE)
• GPQA
• MMLU
• MMMU
• AIME (American Invitational Mathematics Examination)
• Humanity's Last Exam
• Claude Opus 4
• Claude Sonnet 4
• Claude 4
• GPT-5
• GPT-4o
• Large Language Model
• AI safety
• Chatbot Arena

Topics:AI ModelsGoogle DeepMindLarge Language Models

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Google DeepMind. "Gemini 2.5: Our newest Gemini model with thinking" (March 25, 2025). blog.google/...i-model-thinking-updates-march-2025
2. ^1 ^2 ^3 ^4 ^5 ^6Google DeepMind. "Google I/O 2025: Updates to Gemini 2.5 from Google DeepMind" (May 2025). blog.google/...google-gemini-updates-io-2025
3. ^1 ^2 ^3 ^4 ^5 ^6Google Cloud. "Gemini 2.5 Updates: Flash/Pro GA, SFT, Flash-Lite on Vertex AI" (June 2025). cloud.google.com/...sh-lite-flash-pro-ga-vertex-ai
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7Google Cloud Documentation. "Gemini 2.5 Pro | Generative AI on Vertex AI". docs.cloud.google.com/...2-5-pro
5. ^1 ^2 ^3Google AI for Developers. "Gemini Developer API Pricing". ai.google.dev/...pricing
6. ^1 ^2 ^3 ^4 ^5 ^6Google Developers Blog. "Gemini 2.5 Pro Preview: even better coding performance" (May 6, 2025). developers.googleblog.com/...ed-coding-performance
7. ^1 ^2 ^3 ^4 ^5Google DeepMind. "Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context" (technical report). storage.googleapis.com/...gemini_v2_5_report.pdf
8. Helicone. "Gemini 2.5 Pro: Benchmarks & Integration Guide for Developers". helicone.ai/...gemini-2.5-full-developer-guide
9. ^Google AI for Developers. "Models | Gemini API". ai.google.dev/...models
10. Google Cloud. "Expanding Gemini 2.5 Flash and Pro capabilities". cloud.google.com/...2-5-flash-and-pro-capabilities
11. ^Google Developers Blog. "Gemini 2.0 model updates: 2.0 Flash, Flash-Lite, Pro Experimental" (February 2025). blog.google/...gemini-model-updates-february-2025
12. ^1 ^2Google AI Developers Forum. "Gemini 2.5 Pro's Response Quality Regression". discuss.ai.google.dev/...91240
13. ^1 ^2Simon Willison. "Gemini 2.5 Pro" (March 25, 2025). simonwillison.net/...gemini
14. Simon Willison. Blog posts for May 20, 2025 (Google I/O coverage). simonwillison.net/...20
15. ^1 ^2Vellum AI. "LLM Leaderboard: Reasoning models". vellum.ai/llm-leaderboard
16. ^1 ^2 ^3 ^4Artificial Analysis. "Gemini 2.5 Pro: Intelligence, Performance & Price Analysis". artificialanalysis.ai/...gemini-2-5-pro
17. ^1 ^2METR. "Time horizons" (METR-Horizon-v1.0 measurements, including Gemini 2.5 Pro Preview). metr.org/time-horizons
18. ^LMSYS / Chatbot Arena. "Leaderboard" (snapshots, March-June 2025). lmarena.ai/leaderboard
19. ^1 ^2 ^3Google Developers Blog. "Gemini API I/O updates" (May 2025). developers.googleblog.com/...gemini-api-io-updates
20. ^Anthropic. "Claude 3.7 Sonnet and Claude Code" (February 24, 2025). anthropic.com/...claude-3-7-sonnet
21. ^OpenAI. "Introducing GPT-4.5" (February 27, 2025). openai.com/...introducing-gpt-4-5
22. ^xAI. "Grok 3 Beta: The Age of Reasoning Agents" (February 17, 2025). x.ai/...grok-3
23. ^DeepSeek. "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning" (January 20, 2025). github.com/...DeepSeek-R1
24. Wikipedia. "Gemini (language model)". en.wikipedia.org/...Gemini_(language_model)
25. ^1 ^2 ^3 ^4Google. "Gemini 2.5: Deep Think is now rolling out" (August 1, 2025). blog.google/...gemini-2-5-deep-think
26. ^1 ^2 ^3 ^4 ^5 ^6Google DeepMind. "Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad" (July 21, 2025). deepmind.google/...rnational-mathematical-olympiad
27. ^1 ^2 ^3 ^4Google DeepMind. "Gemini achieves gold-medal level at the International Collegiate Programming Contest World Finals" (September 2025). deepmind.google/...rogramming-contest-world-finals
28. ^Alphabet. "Q1 2026 Earnings: Gemini and Google Cloud Statistics" (April 2026). abc.xyz/investor
29. ^1 ^2 ^3Google AI for Developers. "Gemini Model Deprecations". ai.google.dev/...deprecations
30. ^Google AI for Developers. "Gemini 2.5 Pro text-to-speech" (model page, last updated August 18, 2026). ai.google.dev/...gemini-2.5-pro-preview-tts
31. ^1 ^2Google AI for Developers. "Release notes | Gemini API" (May 20, 2025 entry). ai.google.dev/...changelog
32. ^Google. "Gemini logo (Google Image Library)". Google Image Library publication permission. Original official raster, unchanged.
33. ^1 ^2 ^3Beebom. "Google Unveils Gemini 2.5 Pro, Shattering Records on Humanity's Last Exam" (March 2025). beebom.com/...tering-records-on-humanity-last-exam
34. ^1 ^2 ^3 ^4 ^5Analytics Vidhya. "Gemini 2.5 Pro is Now #1 on Chatbot Arena with Impressive Jump" (March 2025). analyticsvidhya.com/...o-is-now-1-on-chatbot-arena
35. ^Google. "Gemini 2.5 Pro: Access Google's latest preview AI model" (June 5, 2025). blog.google/...gemini-2-5-pro-latest-preview
36. ^1 ^2Google DeepMind. "Gemini Pro" model page with Gemini 2.5 Pro Experimental benchmark table (archived April 15, 2025). web.archive.org/...pro
37. ^Google AI for Developers. "Gemini thinking" (archived July 31, 2025). web.archive.org/...thinking

Show all 37 references