Every Major AI Model Expected in May 2026 — What's Confirmed, What's Rumoured, and What It All Means

Why May 2026 Is the Month to Watch

April 2026 was the most consequential month in the history of frontier AI releases. In a single week — April 20 to 24 — OpenAI shipped GPT-5.5, DeepSeek dropped V4 Preview, Moonshot AI launched Kimi K2.6, and xAI released Grok 4.3. Claude Opus 4.7 had landed the week before on April 16. Gemini 3.1 Pro, Llama 4, Qwen 3, and Gemma 4 all shipped within the same six-week window.

May enters with a pipeline that is, if anything, more loaded. Several of the models below were delayed specifically because they were not ready to compete with April's releases. Now they have to ship — or fall further behind. Here is the full status on every model worth tracking.

Model-by-Model Breakdown

1. Claude Mythos (Anthropic) — The One Everyone Is Watching

Status: Restricted preview. ~50 partner organizations. No public release date confirmed.

Claude Mythos is Anthropic's next-generation flagship — confirmed to exist after a data leak on March 26, 2026. Internal documents described it as "by far the most powerful AI model we have ever developed," positioned above the existing Opus tier in a new model class codenamed Capybara. Anthropic confirmed training is complete and that it is being trialed with early access partners, with cybersecurity organizations prioritized first under Project Glasswing.

Gated evaluations show 93.9% on SWE-bench Verified and 94.6% on GPQA Diamond — benchmark numbers that would, if confirmed publicly, reset performance expectations across the industry. For context, Claude Opus 4.7 currently leads coding benchmarks at 64.3% on SWE-Bench Pro. Mythos appears to be in a different category entirely.

Anthropic's stated reason for the restricted rollout is the model's cybersecurity capabilities — its ability to identify software weaknesses at a level that demands careful access controls before broader release. Only ~50 organizations currently have access, compared to the thousands in OpenAI's Trusted Access for Cyber program for GPT-5.5-Cyber.

Why it matters: If Mythos goes public in May, it immediately becomes the strongest argument for Claude over GPT-5.5 in coding and reasoning. It would also force OpenAI to accelerate whatever comes after GPT-5.5.
Probability of May drop: Possible, not confirmed. Anthropic has given no public timeline.

2. Meta Avocado — The Delayed Wildcard

Status: Expected May or June 2026. Delayed from March. Performance concerns flagged internally.

Meta's next-generation frontier model, codenamed Avocado, was originally targeted for late 2025, slipped to March 2026, and is now expected to land "May or June" according to Reuters sources. The reason for multiple delays is documented: internal tests showed Avocado performing between Google's Gemini 2.5 and Gemini 3.0 — incremental progress, but not the leap needed to compete with GPT-5.5 or Claude Opus 4.7 on the benchmarks that matter to developers.

The situation is unusual enough that Meta's AI leadership reportedly discussed temporarily licensing Google's Gemini technology to fill the gap — though no decision has been made. Meta is planning to spend between $115 billion and $135 billion on AI capital expenditure in 2026, so resources are not the constraint. Execution is.

Avocado is expected to be a proprietary model, a departure from Meta's open-source Llama strategy. That shift raises questions about Meta's long-term positioning in the open versus closed model debate.

Why it matters: Meta has 3+ billion users across its platforms. Even a mid-tier frontier model embedded into WhatsApp, Instagram, and Meta AI represents enormous distribution leverage. The model does not need to beat GPT-5.5 to matter at scale.
Probability of May drop: Plausible — Reuters said May or June. Performance pressure is building.

3. DeepSeek V4 Full Release

Status: Preview live since April 24. Full open-weight release expected imminently.

DeepSeek V4 Preview — both Flash and Pro variants — dropped on April 24, 2026, exactly 24 hours after GPT-5.5. The timing was not coincidental. V4 Pro is a 1.6 trillion parameter mixture-of-experts model (49 billion active parameters), the largest open-weight model ever released publicly. V4 Flash runs at 284 billion parameters with 13 billion active.

The pricing is the story. V4 Flash costs $0.14 per million input tokens and $0.28 per million output — undercutting every other frontier-class model including GPT-5.4 Nano, Gemini 3.1 Flash, and Claude Haiku 4.5. V4 Pro at $3.48 per million output tokens undercuts GPT-5.5 ($30) and Claude Opus 4.7 ($25) by a factor of roughly 7x to 9x.

The benchmark reality is more measured: DeepSeek's own technical report acknowledges V4 "falls marginally short of GPT-5.4 and Gemini 3.1 Pro, suggesting a developmental trajectory that trails state-of-the-art frontier models by approximately three to six months." Coding competition benchmarks are "comparable to GPT-5.4." Both V4 models are text-only — no audio, images, or video.

DeepSeek is also reportedly seeking a $20 billion valuation in a fundraising round from Tencent and Alibaba, primarily to retain AI researchers being poached by rival labs.

Why it matters: Cost efficiency at this scale reshapes developer economics. For high-volume API use cases where multimodal is not required, V4 is the cheapest frontier-class option available. The full open-weight release will allow fine-tuning and self-hosting.
Probability of May drop: High. Preview is live. Full release is the expected next step.

4. Nemotron 4 (NVIDIA)

Status: Confirmed upcoming. No release date. Built via Nemotron Coalition with Mistral AI and seven other labs.

NVIDIA announced the Nemotron Coalition at GTC in March 2026 — a group of eight AI companies co-developing open frontier models on NVIDIA DGX Cloud, with outputs feeding directly into the Nemotron 4 model family. Partners include Mistral AI, Black Forest Labs, Cursor, and LangChain. The coalition's first deliverable is a base model co-developed with Mistral, which NVIDIA will open-source on completion.

The current generation — Nemotron 3 — launched at GTC with Ultra, Super, Nano, and VoiceChat variants. Nemotron 3 Nano Omni, released April 29, 2026, added multimodal capabilities (video, audio, image, and text) in a 30B hybrid mixture-of-experts architecture, claiming 9x higher throughput than other open omni models. Nemotron 3 Super and Ultra are expected in the first half of 2026. Nemotron 4 follows after that.

Why it matters: NVIDIA's open model strategy is not altruism — it builds switching costs into its GPU and NIM microservices ecosystem. Nemotron 4 is the most anticipated open frontier model from a hardware-first company, and it is being built collaboratively across the strongest open-source AI labs.
Probability of May drop: Low. Nemotron 3 Super and Ultra ship first. Nemotron 4 is a later-2026 target.

5. GR00T N2 (NVIDIA)

Status: Previewed by Jensen Huang at GTC. Expected "by end of 2026." Not confirmed for May.

GR00T N2 is NVIDIA's next humanoid robot foundation model — a vision language action (VLA) model built for physical AI. Jensen Huang previewed it at GTC in March 2026 and claimed it succeeds at new tasks in new environments more than twice as often as current competing VLA models. It currently tops MolmoSpaces and RoboArena benchmarks for generalist robot policies.

The current production model, GR00T N1.7, is already commercially deployed: LG Electronics and NEURA are using it for humanoid robot scaling, and Johnson & Johnson MedTech and Toyota Research Institute are using the associated Cosmos platform for physical AI training.

GR00T N2 is a different category from the other models on this list — it is not a language or coding model, but a robot action model. Its impact will be felt in manufacturing, logistics, and healthcare robotics rather than in developer tools or consumer AI products.

Why it matters: Physical AI is the next frontier. GR00T N2's "twice as often" success rate claim in novel environments is the kind of step-change that makes humanoid robots practically viable for industrial deployment.
Probability of May drop: Low. NVIDIA said "end of 2026." May is unlikely.

6. GPT-5.5-Cyber Rollout (OpenAI) — Already Happening

Status: Rolling out now. Announced April 30, 2026.

Technically already in motion as May begins. Sam Altman announced on April 30 that GPT-5.5-Cyber — a cyber-permissive variant of GPT-5.5 built for critical infrastructure defenders — is rolling out to verified users via OpenAI's Trusted Access for Cyber program at chatgpt.com/cyber. Government coordination is planned. This is not a new model, but the most significant AI cybersecurity deployment of 2026 so far, and its rollout will dominate the first weeks of May.

Why it matters: The race between OpenAI's TAC program (thousands of defenders) and Anthropic's Mythos Preview (~50 organizations) sets the tone for how frontier AI enters national security and critical infrastructure.

The Models That Already Landed in Late April

For context, these were confirmed before May began and are already in users' hands:

What May's Releases Mean for Developers and Teams

The pace is no longer sustainable for teams that evaluate every new model from scratch. The practical implication of the April–May release window is that multi-model routing is becoming the default architecture — you commit to infrastructure that can swap models, not to any single model.

For coding workflows: Claude Opus 4.7 leads today. If Mythos drops publicly in May, it resets that benchmark immediately. Build on interfaces (Claude Code, Codex, Cursor) rather than hardcoding model dependencies.

For cost-sensitive API use: DeepSeek V4 Flash at $0.14 per million input tokens is the cheapest frontier-class option available. For text-only, high-volume workloads, no competitor is close on price. The full open-weight release in May makes self-hosting viable.

For enterprise and government: GPT-5.5-Cyber's TAC rollout is the most significant AI security deployment of 2026. If your organization defends critical infrastructure, the access application at chatgpt.com/cyber is the first action item for May.

FAQ

When is Claude Mythos releasing publicly?

No public release date has been announced. Anthropic has confirmed training is complete and that Mythos is in restricted preview with approximately 50 partner organizations through Project Glasswing, prioritizing cybersecurity use cases. A public release is expected but unconfirmed for May 2026.

What is Meta Avocado and when does it release?

Avocado is Meta's next-generation frontier AI model. It was originally targeted for late 2025, delayed to March 2026, and then delayed again to "May or June 2026" according to Reuters. Internal tests showed it performing between Gemini 2.5 and Gemini 3.0 — below expectations. Meta is reportedly considering licensing Google's Gemini technology as a stopgap while Avocado is refined.

Is DeepSeek V4 better than GPT-5.5?

No, by DeepSeek's own admission. The company's technical report states V4 "falls marginally short of GPT-5.4 and Gemini 3.1 Pro, suggesting a developmental trajectory that trails state-of-the-art frontier models by approximately three to six months." On coding benchmarks, V4 is "comparable to GPT-5.4" — one generation behind GPT-5.5. However, V4 is dramatically cheaper: V4 Pro costs $3.48 per million output tokens vs GPT-5.5's $30.

What is Nemotron 4 and how is it different from Nemotron 3?

Nemotron 4 is the upcoming next generation of NVIDIA's open language model family, being built collaboratively through the Nemotron Coalition — a group of eight AI labs including Mistral AI, Cursor, and LangChain. Nemotron 3 (current generation) includes Ultra, Super, Nano, and VoiceChat variants. Nemotron 4 will build on the coalition's jointly trained base model and is expected later in 2026, not in May.

What is GR00T N2?

GR00T N2 is NVIDIA's next humanoid robot foundation model — a vision language action model that Jensen Huang claims succeeds at new tasks in new environments more than twice as often as current competing models. It is not a language model for chat or coding; it is a physical AI model for robotics. NVIDIA previewed it at GTC in March 2026 and expects to ship it by end of 2026. A May release is unlikely.

Which May 2026 AI release will have the biggest impact?

Claude Mythos, if it goes public, would have the largest immediate impact — its gated benchmark numbers (93.9% SWE-bench Verified) suggest a step-change in capability that would reset the coding and reasoning leaderboards. Meta Avocado would have the largest distribution impact given Meta's 3+ billion user base, even if the model itself is not frontier-leading. DeepSeek V4's full open-weight release would have the largest impact on developer economics.