Anthropic Launches Claude Opus 5 at Same Price as Predecessor

The new model sets state-of-the-art scores on coding and knowledge-work benchmarks while matching Opus 4.8's pricing and adding stronger alignment and safety controls.

Anthropic released Claude Opus 5 on July 24, 2026, positioning it as a high-capability everyday model that approaches the intelligence of the company’s top-tier Claude Fable 5 at half the price [1].

The practical upshot for developers: Opus 5 costs $5 per million input tokens and $25 per million output tokens — identical to its predecessor, Opus 4.8 — while delivering substantially better performance across coding, knowledge work, and scientific research tasks, according to Anthropic [1].

Benchmark Performance

On Frontier-Bench v0.1, a software engineering evaluation, Opus 5 surpasses all other models tested and more than doubles Opus 4.8’s score at a lower cost per task [1]. On CursorBench 3.2 at maximum effort, it comes within 0.5% of Fable 5’s peak score but at half the cost per task [1].

Anthropic also reports gains in scientific domains. Opus 5 outperforms Opus 4.8 on every life sciences evaluation the company runs, with the largest jumps in organic chemistry — scoring 10.2 percentage points higher on inferring molecular structures from spectroscopy data — and in protein-related tasks, where it scores 7.7 percentage points higher on predicting how sequence variations affect protein function [1].

One area where Opus 5 does not lead: cybersecurity. The model remains behind Anthropic’s Mythos 5 on offensive security tasks, particularly on exploit development [1].

Effort Settings and Speed

Opus 5 introduces a configurable effort setting that lets customers trade off between maximum intelligence and faster, cheaper results by conserving tokens [1]. A Fast mode runs at roughly 2.5 times the default speed and is priced at twice the base rate on the Claude Platform; it is also available through usage credits in Claude Code [1].

Opus 5 is the new default model on Claude Max and the strongest model available on Claude Pro [1].

Alignment and Safety

Anthropic says its automated behavioral audit found Opus 5 to be “our most aligned model to date,” with lower rates of deceptive behavior and greater resistance to misuse attempts than Opus 4.8, Sonnet 5, or Fable 5 [1].

On dual-use risks, the company says Opus 5 does not advance the frontier in biology or offensive cybersecurity, based on evaluations conducted with private-sector and government partners [1]. Although Opus 5 can find vulnerabilities at a rate similar to Mythos 5, it remains substantially behind on exploit development — turning vulnerabilities into active threats [1].

Cyber classifiers on Opus 5 are designed to be around 85% less restrictive than those on Fable 5, allowing source-code vulnerability scanning while blocking binary-based scanning, penetration testing, and exploit generation [1]. Flagged requests in Claude.ai, Claude Code, and Claude Cowork fall back to Opus 4.8 by default [1]. Enterprises and researchers already enrolled in Anthropic’s Cyber Verification Program (CVP) get immediate access to a version of Opus 5 with fewer restrictions [1].

On the biology side, biology-related requests that are blocked on Fable 5 will now route to Opus 5 rather than Opus 4.8, making Opus 5 the most capable generally available model for scientific research under Anthropic’s current safeguard framework [1].

Availability

Claude Opus 5 is available today across all Anthropic platforms via the model identifier claude-opus-5 on the Claude API [1]. Anthropic notes that, consistent with prior Opus models, Opus 5 does not have data retention requirements for general access [1].

Two additional beta updates are shipping alongside Opus 5, though Anthropic did not detail them fully in the announcement [1].


Sources

  1. anthropic.com — Introducing Claude Opus 5

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

Mira Murati's startup ships a multimodal, open-weight model it says balances cost and performance — but acknowledges it isn't the strongest available.

Thinking Machines Lab, the AI startup founded by former OpenAI executive Mira Murati, released its first publicly available model on Wednesday — more than a year after the company was founded [1].

The model, called Inkling, is designed to handle queries across multiple media types while balancing “cost against performance,” the company said [1]. It is open-weight, meaning developers can download and adapt it freely, though the underlying training data and source code are not included [1].

Thinking Machines was candid about where Inkling sits in the competitive field: the company said the model was trained “for strong performance across the board,” but that “it is not the strongest model available today, closed or open” [1].

Filling a Gap in Open-Weight AI

The release comes as the US open-model market is seen as lagging behind Chinese developers in competitive freely available AI products [1]. Meta, previously a prominent player in open model development, has shifted focus toward closed or proprietary offerings, while OpenAI’s portfolio remains dominated by paid products despite releasing a pair of open models last year [1]. Thinking Machines is not aiming to monetize Inkling directly [1].

The startup currently generates revenue through a developer tool called Tinker, which is used for fine-tuning — that is, customizing — AI models [1]. Customers include hedge fund Bridgewater Associates, which uses Tinker to improve AI performance on financial tasks [1].

Backing and Background

Thinking Machines was founded in February 2025 by Murati alongside other former OpenAI employees [1]. The startup raised $2 billion at a $12 billion valuation last year and was reported to be in discussions about a larger follow-on round [1]. Several employees have since departed, moving to companies including Meta and OpenAI [1].

Inkling is also part of a broader product direction Murati has described as “interaction models” — systems built to let people collaborate more naturally with AI [1]. “Our interactions with each other are very rich,” she told Bloomberg last month. “There’s a lot of information in our interactions – when we’re silent, when we’re thinking, when we’re interrupting one another” [1].


Sources

  1. fortune.com — Murati’s Thinking Machines releases first AI model for broad use | Fortune

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

Paid subscribers on Pro, Max, Team, and premium Enterprise seats can use Anthropic's newest model at no extra cost through July 12, 2026, up to half their weekly usage limit.

Anthropic has extended free promotional access to Claude Fable 5 — its newest model — through July 12, 2026 at 11:59 PM Pacific Time, giving paid subscribers more time to test the model without additional charges [1].

The promotion, which originally launched July 1, matters to developers and teams because it lowers the barrier to evaluating a new frontier model mid-cycle, without requiring a separate billing commitment or plan upgrade [1].

Who qualifies and how to access it

Promotional access is available on Pro, Max, Team, and premium seats on seat-based Enterprise plans [1]. Free-plan users are excluded [1]. There is nothing to activate — eligible subscribers can simply select “Fable 5” from the model picker in Claude on the web, Claude Desktop, or Claude Mobile [1].

Developers using Claude Code need version 2.1.170 or later to access Fable 5 [1]. Users of Claude Cowork, Anthropic’s collaborative workspace feature built into Claude Desktop, must be on the latest version of that app [1].

Usage limits and what happens after the promotion

During the promotional period, Fable 5 draws from each subscriber’s existing weekly plan limits, not a separate bucket [1]. Subscribers can use up to 50% of their weekly limit on Fable 5 at no extra cost [1]. Importantly, usage of other Claude models counts against the same overall weekly cap, so the 50% Fable 5 allowance is a ceiling, not an addition [1].

Once a user hits that Fable 5 sub-limit, they have two options: purchase usage credits — billed separately from the subscription — to keep using Fable 5, or switch to another model and continue within their remaining plan limits [1]. If usage credits are already enabled and a user keeps prompting Fable 5 after hitting the limit, billing switches to credits automatically [1].

After July 12, Fable 5 is no longer included in any plan’s weekly limits, and continued access requires usage credits across all plan types [1].

Enterprise seat distinctions

The promotion treats Enterprise seat types differently. Premium seats include Fable 5 access through July 12 at no extra cost, drawing from each member’s seat usage [1]. Standard seats on legacy seat-based Enterprise plans do not include promotional access; all Fable 5 usage on those seats draws from usage credits, meaning the model is unavailable entirely if the organization has not enabled usage credits [1].

API access to Fable 5 is billed separately at standard rates regardless of plan type and is not covered by the promotion [1].

In Claude Code, administrators can restrict Fable 5 availability through managed settings, so users who do not see the model in that environment may have had access blocked by their organization [1].


Sources

  1. support.claude.com — Claude Fable 5 promotional access | Claude Help Center

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

After an 18-day suspension triggered by a US government export control order, Claude Fable 5 is back online globally — with a new safety classifier targeting the bypass technique that prompted the shutdown.

Anthropic has restored global access to Claude Fable 5, its latest general-purpose model, following the lifting of US export controls that had forced a full suspension of the model since mid-June. [1]

The US government applied export controls to both Claude Fable 5 and Claude Mythos 5 on June 12, requiring Anthropic to restrict access to foreign nationals. [1] Because the order took effect immediately and Anthropic had no reliable way to verify nationality in real time, the company suspended access for all users rather than attempt partial enforcement. [1]

The export controls were lifted on June 30, and Fable 5 became available on July 1 to users globally across the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. [1] For Pro, Max, Team, and select Enterprise plan subscribers, Fable 5 is included for up to 50% of weekly usage limits through July 7, after which it will be available via usage credits. [1] Anthropic said it will re-enable access on AWS, Google Cloud, and Microsoft Foundry “as quickly as possible.” [1]

The export control directive was triggered after Amazon researchers published a report describing a method of bypassing Fable 5’s safeguards — specifically, a prompting technique that caused the model to identify software vulnerabilities and, in one case, produce code demonstrating how a vulnerability could be exploited. [1]

Anthropic’s own testing found that the behavior was not unique to Fable 5. The company said several less capable models — including Claude Opus 4.8, GPT-5.5, and Kimi K2.7 — could identify the same vulnerabilities. [1] For the single exploitation demonstration, Anthropic said every model it tested produced the same output, including Claude Haiku 4.5, Sonnet 4.6, Opus 4.6, Opus 4.7, Opus 4.8, GPT-5.4, GPT-5.5, and Kimi K2.7. [1]

Anthropic also concluded that the reported technique “did not expose any unique Mythos-level cyber capabilities” and characterized the bypassed behavior as “routine defensive cybersecurity work.” [1] Nonetheless, the company trained an improved safety classifier to block the specific technique described in the Amazon report. [1] Users whose requests are blocked will be notified, and those requests will be rerouted to Claude Opus 4.8. [1]

The new classifier blocks the reported technique in over 99% of cases, though Anthropic acknowledged that in a very small fraction of cases the model may still provide information — though not at a level of detail useful to a cyberattacker. [1] Researchers from the US Department of Commerce’s Center for AI Standards and Innovation (CAISI) tested both the prior and updated safeguards and, according to Anthropic, agreed that they are “extraordinarily strong.” [1]

Anthropic cautioned that the new classifier comes with a trade-off: it flags benign requests more often during routine coding and debugging tasks. [1] The company said it will continue refining the classifier to reduce false positives. [1]

Fable 5 and Mythos 5 share the same underlying model but were released with different safeguard profiles. [1] Fable 5 launched on June 9 with what Anthropic described as the strongest safeguards it has ever applied to a model, including a deliberately enlarged “safety margin” that causes the model to block some requests that are likely benign in order to make it harder to reach genuinely harmful behaviors. [1] Mythos 5, which carries fewer safeguards, was released only to a small number of trusted partners in a program called Glasswing for use in defensive cybersecurity. [1]

Anthropic described Mythos 5 as capable of finding and exploiting software vulnerabilities “more effectively than any other model — and all but the most skilled human security experts,” making it a high-value target for potential misuse. [1] Access to Mythos 5 has been restored for a set of US organizations following government approval on June 26, and Anthropic said it is continuing to coordinate with the government to expand access to additional domestic and international Glasswing partners. [1]

On the broader question of jailbreak resilience, Anthropic said it considers it “probably impossible to make any AI model fully robust” to jailbreaks, and expects that jailbreaks of varying severity will continue to be discovered. [1] The company said no universal jailbreak for Fable 5 has been found at the time of writing, and that expert safety researchers continue to red-team the model. [1]


Sources

  1. anthropic.com — Redeploying Claude Fable 5

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

The new mid-tier model closes the gap with Anthropic's Opus-class models on coding and tool use, launching at introductory rates before a price increase in September.

Anthropic on June 30 released Claude Sonnet 5, a mid-tier model the company says is its most capable Sonnet-class model for agentic tasks — work that involves autonomous planning, tool use, and multi-step execution without constant human input. [1]

The model is positioned between Anthropic’s previous Sonnet-class releases and its higher-end Opus models. According to Anthropic, Sonnet 5’s performance is close to that of Opus 4.8 on agentic benchmarks, while carrying a lower price tag. [1]

On two agentic evaluations — BrowseComp, which tests web search and research, and OSWorld-Verified, which tests computer use — Sonnet 5 is described as a strict improvement over its predecessor, Sonnet 4.6, though Opus 4.8 still leads at higher accuracy levels. [1]

Anthropic says early access partners reported that the model completes complex tasks where previous Sonnet models would stop short, and checks its own output without being explicitly prompted to do so. [1]

Pricing and Availability

Sonnet 5 is available immediately across all Anthropic plans — Free, Pro, Max, Team, and Enterprise — and is now the default model for Free and Pro users. [1] It is also available in Claude Code and through the Claude application programming interface (API) using the model identifier claude-sonnet-5. [1]

Through August 31, 2026, the model is priced at $2 per million input tokens and $10 per million output tokens. [1] After that date, pricing moves to $3 per million input tokens and $15 per million output tokens. [1]

Anthropic notes that Sonnet 5 uses an updated tokenizer — the component that converts text into the numerical units a model processes — which can increase token counts for the same input by roughly 1.0 to 1.35 times depending on content type. [1] The company says the introductory pricing is set to make the transition from Sonnet 4.6 roughly cost-neutral. [1]

Anthropic has also increased rate limits across its Chat, Cowork, Claude Code, and Claude Platform products to accommodate higher token usage at elevated effort levels. [1]

Safety Profile

Anthropic’s pre-deployment safety evaluations found that Sonnet 5 shows a lower overall rate of undesirable behaviors than Sonnet 4.6, including better resistance to malicious requests and prompt injection attacks — attempts by malicious content in the environment to hijack a model’s actions. [1] The model also shows lower rates of hallucination and sycophancy than its predecessor. [1]

However, Anthropic notes that Sonnet 5 scored somewhat higher on misaligned behavior than the more capable Opus 4.8 and Claude Mythos Preview models. [1]

On cybersecurity risk specifically, Anthropic says it did not deliberately train Sonnet 5 on cybersecurity tasks. [1] On evaluations testing the ability to develop software exploits — including one focused on Firefox browser vulnerabilities — Sonnet 5 was never able to produce a full working exploit, though it showed a slightly higher rate of partial success than Sonnet 4.6. [1] Anthropic attributes that marginal increase to general intelligence improvements rather than targeted training. [1]

Because of this, Anthropic has launched Sonnet 5 with cyber safeguards enabled by default, the same safeguards present in Opus 4.7 and 4.8. [1] The company says these safeguards are less strict than those applied to Fable 5, which block a wider range of cybersecurity tasks. [1] Sonnet 5 is part of Anthropic’s Cyber Verification Program, currently available on the native Claude Platform, the Claude Platform on Amazon Web Services, and Claude in Microsoft Foundry; Google Vertex support is listed as coming soon. [1]

Full evaluation details are available in the Claude Sonnet 5 System Card. [1]


Sources

  1. anthropic.com — Introducing Claude Sonnet 5

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

The model, which spent two months atop OpenRouter under a stealth identity, is now publicly available under an MIT license with aggressive promotional pricing.

Chinese food-delivery and super-app company Meituan has officially released LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts (MoE) model optimized for autonomous software engineering, publishing it on GitHub, Hugging Face, and its own platform under a commercially permissive MIT license. [1]

The release also revealed that LongCat-2.0 was the model behind “Owl Alpha,” an anonymous stealth model that had been running on the OpenRouter model-routing platform for approximately two months before Meituan claimed it. [1]

During that unbranded period, Owl Alpha processed roughly 10.1 trillion monthly tokens — averaging 559 billion tokens per day — representing a 242% month-over-month increase in volume that placed it among OpenRouter’s global top three models. [1] By the time Meituan identified itself as the developer, the model had reached the top ranking on the Hermes Agent workspace, second place on Claude Code deployments, and third place across international OpenClaw environments. [1]

Architecture and Context Window

LongCat-2.0 scales to 1.6 trillion total parameters while limiting active computation to an average of 48 billion parameters per token, with dynamic activation ranging from 33 billion to 56 billion parameters depending on query complexity. [1] The model supports a native 1-million-token context window. [1]

To sustain that context window without severe memory overhead, Meituan developed LongCat Sparse Attention (LSA), described as an evolution of DeepSeek Sparse Attention. [1] LSA addresses the quadratic memory costs typical of fine-grained sparse attention through three mechanisms: Streaming-aware Indexing, which restructures token selection into sequential memory blocks to improve High Bandwidth Memory (HBM) utilization; Cross-Layer Indexing, which amortizes attention calculation costs across adjacent model layers; and Hierarchical Indexing, which applies a two-stage coarse-to-fine scoring approach to filter token candidates before fine-grained selection. [1]

The architecture also incorporates an N-gram Embedding module that appends 135 billion parameters to a 5-gram token combination framework, expanding the core embedding space by roughly 100-fold to capture local token relationships and reduce memory input/output bottlenecks during large-batch inference. [1]

Post-Training and Benchmarks

LongCat-2.0’s post-training uses a framework Meituan calls Multi-Teacher Optimization via Mixture of Specialized Experts (MOPD), which separates optimization into three independent expert clusters: Agent Experts focused on tool invocation and self-correcting execution loops; Reasoning Experts optimized for multi-hop logic and STEM problem-solving; and Interaction Experts focused on instruction-following, factual grounding, and safety. [1] A dynamic gate-routing mechanism combines these behaviors at runtime. [1]

On the SWE-bench Pro software engineering benchmark, LongCat-2.0 scores 59.5, compared to 58.6 for OpenAI’s GPT-5.5. [1] The model also scores 70.8 on Terminal-Bench 2.1, 77.3 on SWE-bench Multilingual, and 73.2 on the general corporate workflow simulator FORTE. [1] The source notes that LongCat-2.0 trails premium frontier systems such as Claude Opus 4.8 on broader general-agent benchmarks like FORTE and BrowseComp. [1]

Pricing and Commercial Model

Standard pay-as-you-go API pricing is set at $0.75 per million input tokens and $2.95 per million output tokens. [1] A limited-time promotional discount reduces those rates to $0.30 per million input tokens and $1.20 per million output tokens. [1] All context-cache hits are processed at no charge under either pricing tier. [1]

Meituan also offers structured “Token Packs” — fixed volumetric allocations valid for 30 days — released via flash sales four times daily at 10:00, 16:00, 21:00, and 23:00 Beijing Time on a first-come, first-served basis. [1] The zero-cost cache-hit policy means that in long agentic sessions where a model repeatedly references the same large codebase, developers are only billed for cache misses and new output tokens. [1]

Trained Entirely on Domestic Chinese Chips

Meituan states that LongCat-2.0 was trained on a cluster of more than 50,000 domestic Chinese Application-Specific Integrated Circuits (ASICs), without relying on Nvidia GPUs. [1] The source frames this as evidence that near-frontier models can be scaled on non-U.S. silicon, though it also notes the broader context: the U.S. government has pressured American AI labs to restrict access to their latest models, with OpenAI limiting access to its GPT-5.6 models and Anthropic taking its Claude Fable 5 / Mythos 5 models entirely offline following government orders. [1]

Licensing and Company Background

The MIT license allows developers to modify, integrate, and redistribute LongCat-2.0 within closed-source commercial products without any obligation to disclose derivative work, in contrast to copyleft licenses such as the GNU General Public License (GPL). [1]

Meituan was founded in March 2010 by Wang Xing and reports more than 770 million annual transacting users and a merchant network of over 14.5 million. [1] The company trades on the Hong Kong Stock Exchange. [1] Its AI push began in late 2025 with LongCat-Flash, a 560-billion-parameter MoE model, followed by the reasoning-focused LongCat-Flash-Thinking. [1]


Sources

  1. VentureBeat — Meituan open sources LongCat-2.0, the 1.6T, near-frontier agentic coding model that's been leading OpenRouter — trained entirely on Chinese chips

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

The new model matches pricier 'Thinking' models on medical benchmarks, is free to all users, and was validated by a network of more than 260 physicians.

OpenAI has upgraded ChatGPT’s healthcare capabilities by deploying a new model called GPT-5.5 Instant, the company announced this week. [1]

According to OpenAI, GPT-5.5 Instant matches the performance of the company’s more expensive “Thinking” models — its reasoning-focused tier — on machine-evaluated health benchmarks including HealthBench and HealthBench Professional, but at a significantly lower cost. [1]

The model is available to all ChatGPT users, including those on the free tier, though usage limits apply. [1]

When evaluated against physician-written answers, GPT-5.5 Instant’s responses scored higher on accuracy, clarity, and completeness, OpenAI said. [1] The company also reported that the rate of incorrect health statements produced by the model has dropped 71 percent over the past two months. [1]

The improvements were informed by a physician review network of more than 260 doctors spanning 60 countries, who collectively reviewed more than 700,000 model responses. [1]

OpenAI says more than 230 million people use ChatGPT weekly for health-related queries, citing use cases such as interpreting lab results, preparing for medical appointments, and navigating insurance questions. [1] The company also offers separate products aimed at medical professionals, including ChatGPT for Clinicians and OpenAI for Healthcare. [1]


Sources

  1. The Decoder — ChatGPT's new health upgrade beats doctor-written answers, OpenAI says

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

Google's new Flash model claims frontier-level coding and agentic performance at four times the speed of competing models and less than half the cost.

Google has released Gemini 3.5 Flash, the first model in its new Gemini 3.5 family, positioning it as a high-speed, low-cost option for developers building agentic and coding applications. [1] A companion model, Gemini 3.5 Pro, is already in internal use and is scheduled to roll out next month. [1]

The company says 3.5 Flash outperforms its predecessor, Gemini 3.1 Pro, on several agentic and coding benchmarks, including Terminal-Bench 2.1 (76.2%), GDPval-AA (1,656 Elo), and MCP Atlas (83.6%), and leads on multimodal understanding with a score of 84.2% on the CharXiv Reasoning benchmark. [1] Google also claims the model generates output tokens at four times the speed of other frontier models. [1]

Pricing is positioned as a competitive differentiator: Google states that 3.5 Flash can complete complex tasks at less than half the cost of other frontier models. [1] The model lands in the top-right quadrant of the Artificial Analysis index, a third-party benchmark that tracks the relationship between model quality and speed. [1]

For developers, 3.5 Flash is generally available through the Gemini API in Google AI Studio and Android Studio, the Google Antigravity orchestration harness, the Gemini Enterprise Agent Platform, and Gemini Enterprise. [1] It is also now the default model powering the Gemini app and AI Mode in Google Search globally. [1]

A key capability highlighted by Google is the model’s integration with its Antigravity harness, which enables deployment of collaborative subagents for multi-step, long-horizon workflows. [1] Google describes use cases ranging from automatically renaming and categorizing unstructured assets to synthesizing research papers and generating playable games. [1]

Several enterprise partners are already piloting the model. Shopify is using parallel subagents to analyze data for merchant growth forecasting. [1] Macquarie Bank is testing the model’s ability to reason over documents exceeding 100 pages to accelerate customer onboarding. [1] Salesforce is integrating 3.5 Flash into its Agentforce platform to automate enterprise tasks using multi-turn tool calling. [1] Xero is deploying agents to manage multi-week workflows such as gathering supplier information for 1099 tax forms. [1] Databricks is applying agentic workflows to monitor real-time data, diagnose issues, and propose fixes for data scientists. [1] Fintech company Ramp is using the model’s multimodal understanding to improve optical character recognition (OCR) on complex invoices. [1]

On the consumer side, Google is introducing Gemini Spark, a personal AI agent powered by 3.5 Flash that is described as running continuously to take actions on a user’s behalf. [1] Spark is beginning to roll out to trusted testers, with a beta planned for Google AI Ultra subscribers in the United States the following week. [1]

Google says Gemini 3.5 was developed under its Frontier Safety Framework, with strengthened safeguards against generating harmful content in cybersecurity and chemical, biological, radiological, and nuclear (CBRN) domains. [1] The company also notes the use of interpretability tools designed to examine the model’s internal reasoning before it produces a response. [1]


Sources

  1. Google AI Blog — Gemini 3.5: frontier intelligence with action

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

Google DeepMind's new open-weight text diffusion model generates up to 1,000 tokens per second on a single H100, making it roughly four times faster than comparable autoregressive Gemma models for local, single-user workloads.

Google DeepMind has released DiffusionGemma, an experimental open-weight model that generates text through a diffusion process rather than the token-by-token approach used by most large language models (LLMs), and is making the weights available immediately on Hugging Face under an Apache 2.0 license. [1,2]

Unlike autoregressive models, which produce text left to right one token at a time, DiffusionGemma starts with a block of 256 random placeholder tokens and refines them across multiple passes until coherent text emerges — a technique borrowed from image generation AI. [1,2] The model finalizes all outputs in one large block, a process Google describes as “denoising” a text canvas. [1]

DiffusionGemma uses a Mixture of Experts (MoE) architecture — a design in which several specialized sub-networks exist side by side and only the relevant ones activate for a given input — with 26 billion total parameters but only 3.8 billion activated during inference. [1,2] When quantized to lower numerical precision, the model fits within 18 GB of video RAM (VRAM), making it compatible with high-end consumer GPUs. [1,2]

In benchmarks, DiffusionGemma reaches approximately 700 tokens per second on an Nvidia RTX 5090 and more than 1,000 tokens per second on a single Nvidia H100 AI accelerator. [1,2] Nvidia also reports around 150 tokens per second on the DGX Spark deskside system and up to 800 tokens per second on the DGX Station. [2] Those figures represent roughly four times the throughput of similarly sized autoregressive Gemma models in single-user local inference. [1,2]

The speed advantage stems from a shift in hardware bottleneck. With autoregressive models, single-user inference is typically constrained by memory bandwidth, leaving GPU compute units largely idle while waiting for data. [2] By processing up to 256 tokens in parallel, DiffusionGemma moves the bottleneck toward raw compute, keeping GPUs more consistently occupied. [1,2]

Google and Nvidia note that the advantage is specific to local, single-user scenarios. In cloud deployments serving many parallel requests, autoregressive models already keep hardware busy, and DiffusionGemma can actually increase costs in that setting, according to Google. [2] Google also notes that on shared-memory systems such as Apple Silicon, which face their own memory-bandwidth constraints, the speed gap over autoregressive models is likely smaller. [2]

The parallel generation approach also opens up task categories where autoregressive models struggle. Because every token can reference every other token in the block — including tokens that appear later in the sequence — DiffusionGemma is better suited to tasks such as inline text editing, filling gaps in code, molecular amino acid sequencing, and mathematical graphing. [1,2] Google highlights Sudoku solving as a demonstration: each cell depends on future cells, a dependency structure that trips up standard autoregressive models but is handled more naturally by DiffusionGemma’s continuous self-correction over the full token block. [1,2]

The tradeoff is output quality. Diffusion-based text generation carries a higher error rate than autoregressive generation; a single badly predicted token can render an entire block meaningless and require regeneration, whereas a single bad pixel in an image diffusion model rarely ruins the output. [1] Diffusion models also consume more resources when the desired output is short, since the parallel denoising process does significant work even to produce just a few tokens. [1] Google explicitly recommends its standard Gemma 4 models when quality is the priority, positioning DiffusionGemma as a tool for researchers and developers experimenting with fast local workflows. [2]

Google worked with Nvidia on optimization. Nvidia quantized the model for the RTX 5090 and RTX 4090 and optimized it for Hopper and Blackwell server architectures. [2] The model is also available through the Gemini Enterprise Agent Platform Model Garden and Nvidia NIM. [2]

DiffusionGemma works out of the box with Hugging Face Transformers, vLLM (with Red Hat integration support), and MLX. [2] For fine-tuning, Google points to its own JAX-based Hackable Diffusion toolkit, along with Unsloth and the Nvidia NeMo Framework. [2] Support for llama.cpp is planned. [2]

The release builds on earlier internal work: Google DeepMind had previously demonstrated Gemini Diffusion, an experimental text diffusion model that cited speeds of 1,479 tokens per second and performed roughly on par with Gemini 2.0 Flash-Lite in benchmarks. [2] DiffusionGemma is the first member of the open Gemma 4 family to use the diffusion approach. [1] The startup Inception has been pursuing a similar direction; its Mercury 2 model shipped in early 2026 as what the company describes as the first diffusion-based reasoning model. [2]


Sources

  1. Ars Technica — Google's latest DiffusionGemma open AI model comes with a 4x speed boost
  2. The Decoder — Google's new open model DiffusionGemma generates text from noise instead of word by word

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.