AI labs continue rolling out new models at a rapid pace. While each release claims improvements in speed and performance over earlier versions, not every update represents a true leap forward, regardless of company announcements. Real strengths become clear only when models are compared side by side: where rivals fall short or stand out, which systems offer unique skills, and which simply match current benchmarks.

Our Model Release Tracker clarifies how each model compares to others and whether it merits further exploration. Although we do not evaluate every entry, we highlight essential details and share expert testing results when available. Certain models also receive an Expert Score. Details on our evaluation methods appear in a separate overview.

Below are the major model launches of 2026 to date. The list updates with each significant release.

Update: Fable 5 Access

Fable 5 carries higher usage costs, yet Anthropic offered it without added fees to paid subscribers on Pro, Max, Team, and select Enterprise plans. The company had planned to move Fable 5 to token-based billing on July 7, but later extended free access for paid plans until July 12.

Users may apply up to half of their weekly limit to Claude Fable 5. Beyond that threshold, they can continue with usage credits or switch models to stay within remaining allowances, according to an Anthropic statement on X.

After Fable 5 and Mythos 5 were withdrawn following government guidance shortly after launch, authorities permitted limited Mythos 5 access for select partners on June 26. On June 30 the Department of Commerce removed export restrictions on both models. Anthropic restored worldwide Fable 5 availability on July 1, while Mythos 5 remains restricted to designated institutions.

Sonnet 5

Anthropic | June 30, 2026

What it does: Shifting emphasis from recent Opus releases, Anthropic introduced Sonnet 5. The model can create plans, operate tools such as browsers and terminals, and function independently at levels previously requiring larger systems, the company stated. Performance matches Opus 4.8 from the prior month while carrying lower costs. Pricing begins at $2 per million input tokens and rises to $3 per million in September. Sonnet 5 now serves as the default option for Free and Pro plans and is open to Max, Team, and Enterprise users.

In line with broader industry focus on agent capabilities, Sonnet 5 performs strongly on computer-use and agentic-coding evaluations. Testing showed it handling complex tasks beyond the reach of earlier Sonnet versions. Automatic safeguards accompany the release, reflecting both recent attention around Mythos and general capability gains.

Why it matters: Given the fast pace of development, Sonnet 5 likely marks a meaningful advance over Sonnet 4.6 from February. Release timing also stands out; Anthropic noted reduced capacity for high-risk cybersecurity tasks compared with current Opus models, providing context after the complicated introduction of Fable 5 and Mythos 5. Still, Sonnet 5 recorded a higher rate of misalignment than Mythos Preview.

Fable 5 and Mythos 5

Anthropic | June 9, 2026

What it does: ZDNET contributor David Gewirtz described Fable 5 as a safer, public version of Mythos. Anthropic made Mythos 5 available to users who held Mythos Preview access via Project Glasswing. Fable 5 blocks responses to high-risk topics such as cybersecurity and biological weapons yet retains Mythos-level performance. Mythos 5 was slated for wider rollout through a structured program, but both models were removed within four days at the direction of US officials.

Why it matters: The releases generated attention due to their claimed capabilities and abrupt withdrawal. Fable 5 raised concerns among safety testers when it downgraded to Opus on certain queries, eroding trust. Amazon researchers later bypassed Fable 5 restrictions and notified the White House, though Anthropic described the event as limited in scope.

The episode marks an uncommon public intervention by authorities in a domestic AI product and may signal tighter oversight for advanced models under the current administration.

MAI-Thinking-1

Microsoft AI | June 2, 2026

What it does: Unveiled at the Build conference, this 35-billion-parameter model targets multi-step agent tasks. It achieved comparable results to Opus 4.6 on the SWE Bench Pro coding benchmark. Microsoft emphasized that the model was trained solely on clean, commercially licensed data, addressing growing concerns over copyright litigation.

Why it matters: This marks Microsoft AI's first reasoning model, an important step for the lab. While most competitors already offer multiple advanced reasoning systems, Microsoft positions the release toward enterprise needs.

Claude Opus 4.8

Anthropic | May 28, 2026

What it does: Replacing Opus 4.7 at the same price point, Opus 4.8 delivers faster reasoning modes at one-third the prior cost. The model strengthens coding performance, surpassing 4.7 on two benchmarks though not yet exceeding GPT 5.5. It also records gains in measures of user support and alignment, according to Anthropic.

Why it matters: Anthropic continues to stress safety and transparency. The company reported a 92 percent honesty rate for Opus 4.7 and claims substantially lower misalignment for 4.8, aligning it closer to Mythos Preview standards.

Gemini 3.5 Flash

Google | May 19, 2026

What it does: Launched at Google I/O, the 3.5 family targets agent development. Gemini 3.5 Flash improves on Gemini 3.1 Pro across several coding and agent benchmarks and now powers AI Mode in Search and the Gemini app. Optimized for speed and lower cost, it still manages extended agent tasks. The Pro variant is expected in June.

Why it matters: The model's primary role in Search means it shapes daily user experiences. Accuracy and reduced hallucinations are therefore critical, yet the system card omits specific data on those metrics.

GPT-5.5 Instant

OpenAI | May 5, 2026

What it does: The lighter edition of GPT-5.5 reduces verbosity compared with GPT-5.3 Instant and lowers hallucination rates. OpenAI reported 52.5 percent fewer fabricated claims on high-stakes prompts in medicine, law, and finance.

Why it matters: GPT-5.5 Instant becomes the default in ChatGPT. Fewer hallucinations in a widely used fast-response model could limit misinformation, especially on health-related queries.

(Disclosure: Ziff Davis, ZDNET's parent company, filed an April 2025 lawsuit against OpenAI alleging copyright infringement in training and operating its AI systems.)

Nemotron 3 Nano Omni

Nvidia | April 28, 2026

What it does: The newest open model in Nvidia's Nemotron series enables multimodal agent input. It processes visual, audio, and text data within a single perception-to-action loop, unifying capabilities previously handled by separate systems.

Why it matters: Traditional agent setups require distinct models for different modalities, increasing latency and expense. Nvidia's unified approach could streamline workflows and cut token consumption. The model is available on Hugging Face.

GPT-5.5

OpenAI | April 23, 2026

Expert Score: 93/100

What it does: Tester David Gewirtz assigned GPT-5.5 an A- rating, noting it improves on GPT-5.4 in agentic coding, concept identification, scientific research, and factual accuracy.

Why it matters: Although gains over the immediate predecessor are incremental, the sub-two-month interval between 5.4 and 5.5 underscores accelerating release cycles driven by agentic coding advances. Labs increasingly rely on AI to develop subsequent models.

ChatGPT Images 2

OpenAI | April 23, 2026

What it does: Shortly after retiring Sora, OpenAI released Images 2. Early reviewer David Gewirtz found the model engaging, a clear advance, and practical for professional tasks.

Why it matters: The launch suggests OpenAI still views image generation as relevant to enterprise use, particularly following Anthropic's Claude Design release.