Anthropic’s Opus 4.8 claims lower misalignment, faster coding

Anthropic’s Claude Opus 4.8 arrives May 28, 2026 with faster “thinking modes” at one-third the cost of Opus 4.7, alongside claims of substantially lower misalignment versus the prior version—an update Anthropic also benchmarks against its Mythos Preview work.

By the time Anthropic shipped Claude Opus 4.8 on May 28, 2026, the company was already arguing with itself—in the best sense of the phrase—about what “better” actually means.

Opus 4.8 is described as replacing Opus 4.7 starting that day “at the same price. ” with faster thinking modes for one-third the cost of the earlier version. according to Anthropic. The pitch isn’t just speed. Anthropic says the model prioritizes coding abilities and scores higher than Opus 4.7 on two coding benchmarks. while not fully besting OpenAI’s GPT 5.5.

The harder part of Anthropic’s case sits inside the safety language. In the release. the company says Opus 4.8 “reaches new highs” on its measures of prosocial traits—supporting user autonomy and acting in the user’s best interest—though it admits that the definitions for what those measures mean remain “murky.”.

Then comes the specific claim that matters to anyone who follows AI behavior tests closely: Anthropic says Opus 4.7 had a 92% honesty rate. The company adds that it claims Opus 4.8 shows “substantially” lower rates of misalignment than Opus 4.7. It also compares Opus 4.8’s alignment to Mythos Preview—an earlier. non-public model that Anthropic framed as posing a security threat.

This is where the tension shows, not in a slogan, but in sequencing. Anthropic has long emphasized model safety and interpretability, and Opus 4.8 reads like a continuation of that push. But using a Mythos Preview comparison as a yardstick also highlights how fast the standards—and the framing—are moving in frontier model releases.

The same release cycle is playing out across other major updates listed in 2026. On May 5. 2026. OpenAI rolled out GPT-5.5 Instant. describing it as less verbose than GPT-5.3 Instant and touting fewer hallucinations and improved factuality. OpenAI says GPT‑5.5 Instant produced 52.5% fewer hallucinated claims than GPT‑5.3 Instant on high-stakes prompts covering areas like medicine. law. and finance.

GPT-5.5 Instant also replaces GPT-5.3 as the default model in ChatGPT, meaning the impact isn’t confined to lab benchmarks. The practical question is straightforward: if fewer hallucinations reach everyday prompts. how much less misinformation could make it into routine health questions and other day-to-day use.

OpenAI’s litigation backdrop sits in the disclosure attached to this roundup: Ziff Davis, ZDNET’s parent company, filed an April 2025 lawsuit against OpenAI alleging it infringed Ziff Davis copyrights in training and operating its AI systems.

Nvidia, meanwhile, is pushing the infrastructure side of agentic work. On April 28, 2026, Nvidia released Nemotron 3 Nano Omni, the newest in its open Nemotron family. The model provides agents with multimodal input—letting them “perceive and reason across visual. audio. and textual inputs within a single shared perception‑to‑action loop. ” in Nvidia’s words—unifying multiple capabilities into one system.

The point. as described. is that agent systems often have to juggle separate models for speech. vision. and text. which can slow workflows. undermine the context agents gather. and rack up inference costs. Nvidia’s approach aims to streamline that and reduce token use. The list says to try it on Hugging Face.

Across the OpenAI and Anthropic timelines, coding and agentic delivery keep showing up as the proving ground. On April 23, 2026, OpenAI released GPT-5.5. A ZDNET tester-in-residence. David Gewirtz. gave it an Expert Score of 93/100. and said it technically earned an A- score while being “reductively described as better and faster than GPT-5.4.” In the same assessment. Gewirtz said GPT-5.5 got better at agentic coding. clearly identifying concepts. scientific research. and factual accuracy. The test lost points only for “exuberance.”.

That score matters less for what it celebrates than for what it suggests about release pace. The roundup notes that the quick turnaround from 5.4 to 5.4 is less than two months, and says this indicates how rapidly OpenAI’s agentic coding cycle is accelerating.

OpenAI also released ChatGPT Images 2 on April 23, 2026, after sunsetting Sora—OpenAI’s generative video model and social platform. ZDNET model tester David Gewirtz got an early look at Images 2 and described it as fun. a huge leap. and actually useful for work. The roundup frames the move as part of OpenAI redirecting away from consumer AI products after Sora was discontinued and after Anthropic secured lucrative enterprise contracts. Still. OpenAI’s release of Images 2 suggests it considers image generators relevant enough for enterprise AI—especially on the heels of Anthropic’s Claude Design.

Anthropic’s earlier steps in this chain add more context for why Opus 4.8’s safety claims land the way they do. On April 16, 2026, Anthropic released Claude Opus 4.7. It arrived relatively quickly after Opus 4.6 and. according to the list. boasted new highs in honesty with reduced sycophancy and hallucinations. It also appears to have a cybersecurity knack: the model backed the new Claude Security tool released shortly after the model itself—though the roundup stresses that it is “not Mythos. ” despite what many suspected.

That Claude Security feature is described as scanning your codebase for flaws and helping decide what to fix first.

Before any of these public releases, Anthropic’s Mythos (Preview) added its own kind of pressure. On April 7, 2026, Anthropic introduced Claude Mythos (Preview). The roundup says Mythos is not actually available to the public. but Anthropic still created a media storm by positioning the general-purpose model as too powerful to release as usual. It says the company was alarmed by the security threat it posed. stating that “it is strikingly capable at computer security tasks.”.

In response. Anthropic spearheaded Project Glasswing. a collaborative effort with several rival AI labs—explicitly including Google. Nvidia. and Microsoft—plus security authorities like Palo Alto Networks. The stated goal: to help secure the world’s most critical software. and to prepare the industry for the practices needed to stay ahead of cyberattackers.

The roundup’s stakes are direct: if Anthropic’s guidance is believed—especially the idea that Mythos posed a significant threat so that only a select few partners can access it—then current cybersecurity apparatuses may not be ready for rapidly evolving frontier model capabilities. The list says Mythos, only weeks into its release timeline, has been helping catch software bugs in droves.

Earlier in the year, OpenAI framed GPT-5.4 as built for professional work. GPT-5.4 was released on March 5, 2026, barely three months after GPT-5.2. OpenAI said its testing—stated with a grain of salt until verified by a third party—showed GPT-5.4 matches or outperforms human professionals 83% of the time.

The roundup links that pitch to enterprise trust: models that can handle complex professional tasks with minimal risk. delay. or prohibitively high costs stand a better chance of being taken seriously by companies trying to adopt AI. But it also repeats the claim that GPT-5.4 “clobbers humans on pro-level work in tests – by 83%.”.

And in February, OpenAI and Anthropic both pushed agentic coding further. On Feb. 5. 2026. Anthropic released Claude Opus 4.6. saying it quickly redefined the standard for autonomous agentic work. especially for coding. and improved performance in complex. longer-running tasks. OpenAI’s GPT-5.3-Codex arrived the same day, with OpenAI saying the coding model helped build and debug itself. It can be interrupted and redirected mid-task. is said to have run times of over a day. and is described as having a better grasp on user intent.

The list ends with a reminder that OpenAI was also pushing a speed advantage: it references “Spark model codes 15x faster than GPT-5.3-Codex,” while noting there is a catch.

All of it adds up to one unavoidable reality the tracker tries to capture: newer models don’t always represent a clean leap. Sometimes they’re incremental. But in the specific case of Opus 4.8. Anthropic is asking users to see the difference—through honesty rates. misalignment comparisons. and even through its choice to benchmark against Mythos Preview.

That’s what makes the release feel less like marketing and more like a race you can measure. Safety claims, faster performance, and targeted capability improvements are being stacked release after release—while the question remains whether the industry is getting better faster than it can adapt.

Claude Opus 4.8 Opus 4.7 misalignment honesty rate Anthropic Mythos Preview GPT-5.5 Instant hallucinations ChatGPT Images 2 Nemotron 3 Nano Omni Nvidia agentic coding cybersecurity Project Glasswing Palo Alto Networks Hugging Face