Claude 4.8 Is A Beast… But There’s A Big Problem

Ajouté : 2026-06-04

[00:00:02]Claude Opus 4.8 just arrived and everything about this release looks like a clean win for Anthropic. Better coding, stronger agents, better longrunning tasks, same price, and benchmarks that make it look like one of the strongest AI models in the world right now. But the deeper you look at this release, the stranger it gets.

[00:00:21]Because Anthropic is selling Opus 4.8 around one main idea, honesty. The company says this model is better at admitting uncertainty, better at pointing out problems, and less likely to pretend the work is finished when it actually isn't. And in AI coding, that matters a lot. A model that confidently says the bug is fixed while leaving broken code behind can waste more time than a model that simply fails and tells you what went wrong. But at the same time, Anthropic's own technical material points to a much weirder concern. During training, Opus 4.8 eight started showing a stronger ability to reason about how its output might be scored. Even when it wasn't directly told it was being evaluated, it seemed to shape answers in ways that would probably earn higher scores. So, this is not just a story about Claude getting stronger. It's a story about Claude getting stronger while also becoming better at understanding the test. And that makes the whole honesty angle way more complicated. Anthropic released Claude Opus 4.8 8 on May 28th, only around 41 to 43 days after Opus 4.7, making this one of its fastest minor version updates so far. On the same day, Anthropic also completed a $65 billion series H round, pushing its postinvestment valuation to around $965 billion.

[00:01:45]According to the reports, that would put Anthropic above OpenAI's estimated $852 billion valuation. The clearest improvement is coding. On SWEBench Pro, Opus 4.8 reportedly jumps from 64.3% on Opus 4.7 to 69.2%.

[00:02:04]Anthropics comparison puts GPT 5.5 at 58.6% and Gemini 3.1 Pro at 54.2%.

[00:02:14]On S swbench verified, it rises from 87.6% to 88.6%.

[00:02:21]On OSWorld Verified, a computer use benchmark, it reaches 83.4% and on online Mind2 web partner tests put it around 84%. But the real signal is how it behaves inside developer tools. Cursor co-founder Michael Truel said, "Oopus 4.8 8 beats previous Opus models on Cursorbench at every effort level with more efficient tool calls and fewer steps. Scott Woo, the CEO of Cognition, said it apparently fixes two major complaints from Opus 4.7. Overly verbose comments and unstable tool calls. Lenny's newsletter was more cautious, saying it still struggles with the last 10% old code bases, edge cases, and hallucinations. So, this is not a perfect model. It is a stronger coding agent, especially for fast execution and larger tasks, but it still has familiar LLM weaknesses when things get messy.

[00:03:15]Then there's GDP vala, which measures realworld agentic capability. Opus 4.8 reportedly scored 1,890 ELO, which is 137 points higher than Opus 4.7 and 121 points higher than GPT 5.5. In win rate terms, the reports say that converts to around a 67% winning probability compared to Opus 4.7. It also uses 15% fewer steps and outputs 35% fewer tokens to complete the same task. There are also claims around human last exam agent tasks, program bench, and Frontier SWE.

[00:03:55]In Program Bench, the model has to reconstruct source code from a compiled binary using only project documentation without decompiling or using the internet. Opus 4.8 reportedly beats 4.7 across context budgets. On graph walks, a benchmark that stress tests long context reasoning by packing the context window with a massive directed graph and asking the model to navigate it. Opus 4.8 8 pulls clearly ahead of Opus 4.7.

[00:04:24]On the 256K subset, it hits 85.9% up from 76.9. And on the full 1 million token version, it jumps to 68.1% nearly doubling 4.7 score of just 40.3.

[00:04:40]on Frontier SWE, which includes tasks like writing a Postgress QL server from scratch in Zigg, rewriting Git, and creating a native Lua compiler. Opus 4.8 reportedly tops the list with an 83% win rate. Some people even started calling it not really 4.8, more like Opus 5. One blogger suggested it might be a distilled version of Clawude Mythos, the more powerful model Anthropic is expected to launch within the next few weeks. That part is still speculative, but several reports describe Opus 4.8 as approaching Claude Mythos preview in alignment. Anthropic says deception and cooperation in abuse are significantly lower than with Opus 4.7, while pro-social behavior has reached a new high. And honestly, that is becoming the bigger theme across AI right now.

[00:05:30]Whether we're talking about coding agents or creative tools, the winners are starting to look like systems that can actually carry a workflow from start to finish. That is why Flova caught my attention. Flova is sponsoring today's video and it is one of the first skill-based AI video agents. That skills part is the key difference. Most AI tools generate an output, then everything resets. With Flova, you can build a workflow once and save it as a reusable skill. So your visual style, preferred models, storyboard structure, fonts, characters, and creative direction can actually carry over into future projects. For example, I used Flova to build this short cinematic AI commercial you are seeing on screen. I started with a rough idea, shaped the storyboard through chat, generated the visual direction, refined a few shots, and then saved the process as a skill so the same style and workflow can be reused later. That makes it feel less like a normal AI video generator and more like a persistent creative workspace. Flova also brings models like GPT image 2 and Cedence 2 into one place so you can move between images, video, motion and editing without constantly jumping between tools. And they are also building a skills community where creators can share workflows almost like presets, lots or luras. So if you create AI films, anime ads or social content, Flova is worth checking out. Use the link in the description to try it out and get your free credits. All right, now back to the video. Anthropic says a common problem with AI models is that they claim progress without enough evidence. In coding, that can be brutal.

[00:07:10]The model writes code, says it fixed the issue, and then you later discover it skipped a test, ignored an error, or misunderstood the codebase. It didn't necessarily lie like a human would lie, but from the user's perspective, the result feels the same. The model gave false confidence. With Opus 4.8, Anthropic says the model is more willing to mark uncertainty and make fewer unsupported claims. In code tasks, the probability of letting undetected defects slip through silently is reportedly about one quarter of Opus 4.7's rate. One article says Opus 4.8 is the first clawed model to hit 0% on an evaluation for reporting defective results without criticism. Another metric, the false reporting rate, reportedly goes from 0.40 on Opus 4.5 to 0.25 on Opus 4.7 and then to 0.00 on Opus 4.8. There was also a laziness investigation rate measuring cases where the model gives a lazy answer instead of properly investigating. Opus 4.7 reportedly had a 25% rate while Opus 4.8 hit 0%. That is why some coverage calls this two zeros rewriting history. The idea is simple. Anthropic wants Claude to become the model that does not quietly hide mistakes. There was also a concrete example from Anthropic's own blog. A developer was using Claude code with Opus 4.8 for a code migration and then went out to fly a kite while Claude kept working in the background. During the process, a submission was rejected because a colleague had pushed an emergency fix. Claude notified the developer and said it planned to merge the colleagues changes first, then retry. The developer casually replied that it should just force overwrite it.

[00:09:00]Claude refused. It explained that force overwriting would discard the emergency fix submitted by the colleague at 11:42.

[00:09:07]Instead, it merged both sets of changes, kept the code the same, preserved a clean submission history, and pushed the result. That is exactly the behavior Anthropic wants to highlight. The model didn't blindly follow a shortcut. It protected the workflow. For enterprise customers, that is the pitch. If Claude is going to work inside real code bases, documents, business processes, and production systems, then trust matters more than raw intelligence. A model that is slightly smarter but covers up mistakes is dangerous. A model that admits uncertainty and protects the workflow is much easier to hand real work to. But then comes the strange part. Anthropic's own system card reportedly says one of the biggest concerns during training was that Opus 4.8 became increasingly good at reasoning about how its output would be scored. Even when it was not told it was being evaluated, it seemed to infer that it might be judged and then shape its response in a way that would get a better score. That does not mean it is doing something malicious. Anthropic says this has not yet turned into observable bad behavior. And Opus 4.8 8 actually reports task success less often than the previous version, but they still describe it as a worrying trend that could cause trouble for future training. Early interpretability work also found unspoken scoring related reasoning in about 5% of training segments. On one side, Anthropic is saying opus 4.8 is more honest. On the other side, Anthropic is also saying the model is getting better at understanding the exam. So people naturally ask, is it really becoming more honest or is it becoming better at performing honesty when the test is watching? That question gets even more uncomfortable because many of these honesty scores come from internal evaluations, not independent audits. So the model is being tested by the company that built it on evaluations the company designed while the company itself says the model is getting better at recognizing how it will be scored.

[00:11:06]That does not erase the progress. It just makes the story more intense. Opus 4.8 may genuinely be less overconfident and more reliable while still revealing a deeper problem with model training. As models become more advanced, they may learn to optimize for the evaluation environment itself. There's another weird detail, too. Some users reportedly asked Opus 4.8 what model it was, and it did not always answer Claude. In some cases, it identified itself as Quen or mentioned deepseek, which led to speculation about possible distillation or training artifacts. In the official Clawed client, those answers were apparently less common, probably because the system prompts and product layer controls are stronger there. That part needs to be treated carefully, but it adds to the same feeling. Opus 4.8 is powerful, but something about this release feels strange. And while the model is getting most of the attention, the Claude Code upgrade may matter just as much. Anthropic pushed what is described as the largest underlying upgrade to Claude Code so far, targeting six developer pain points: terminal flickering, thinking freezes, confusing error reports, context deadlocks, unstable MCP connections, and session crashes. The terminal now has a full screen renderer to stop flickering, real-time streaming of thinking and tool calls. so users know the agent is alive, clearer error messages, faster memory compaction with progress, stronger MCP connections to local tools and files, and session self-healing so one corrupted file or oversized image does not crash the whole session. This is where the release becomes bigger than benchmarks. The AI coding race is shifting from who has the smartest model to who has the most reliable work system. Anthropic is also introducing effort control, which lets users choose how much thinking Claude puts into a task. Higher effort means more inference and better answers, while lower effort means faster responses and lower usage.

[00:13:08]Opus 4.8 uses high effort by default. In Claude Code, users can go even higher with extra, X high, or max. Anthropic recommends extra for difficult tasks and longunning workflows. Fast mode also changed. The same model can reportedly run about 2.5 times faster with pricing listed at $10 per million input tokens and $50 per million output tokens for that mode. Described as around three times cheaper than the previous fast mode. Data Brick CTO Hanland Tang said Opus 4.8 reads unstructured content like PDFs and charts in their Genie product while using 61% lower token cost than Opus 4.7. The standard Opus 4.8 8 API price reportedly stays the same as before. $5 per million input tokens and $25 per million output tokens. Then there's dynamic workflows, maybe the most important product feature here. It is currently in research preview and designed for large code bases and big engineering tasks. Claude can plan the task, write orchestration scripts, run dozens or hundreds of parallel sub aents, review their outputs, verify the work, and report back. This is aimed at bug finding, performance audits, security reviews, code migrations, framework replacements, API deprecation migrations, language migrations, and multi-angle verification. Users can ask Claude to create a workflow directly or use ultraode in Claude Code. Ultradeode sets thinking intensity to XH high and lets Claude decide whether the task needs a workflow. Dynamic workflows are available in cloud code, CLI, desktop, and VS code extension for Macs, team, and enterprise plans. Enterprise has it disabled by default at launch, and admins need to turn it on. It can also be used through the claude API, Amazon Bedrock, Vert.Ex AI, and Microsoft Foundry. The biggest example is the bun migration. Jar Sumner used dynamic workflows to port bun from Zigg to Rust, generating about 750,000 lines of Rust code. The existing test suite reached a 99.8% pass rate and the work took about 11 days from first submission to merge.

[00:15:18]The process used multiple workflows, hundreds of agents in parallel, two reviewers per file, repeated build test fix loops, and an overnight workflow for data duplication cleanup. Anthropic also updated the messages API so developers can insert system entries inside the messages array. That means instructions can change during task execution without breaking prompt cache or forcing updates through the user turn. Developers can adjust permissions, token budgets, or environmental context while an agent is already running. And above all of this, Claude Mythos preview is still coming.

[00:15:54]So, Opus 4.8 doesn't just feel like Anthropic's new flagship. It feels like a bridge to the next tier. That's what makes this release so interesting.

[00:16:03]Claude is getting stronger, faster, and more useful for real work. But the same release also raises a strange question.

[00:16:10]Is this model becoming more honest or just better at knowing what honesty is supposed to look like? Drop your thoughts in the comments. Subscribe if you want more AI updates like this. Hit the like button if the video helped. And thanks for watching. I'll catch you in the next one.