Claude Opus 4.8 Field Report: A Correction, Not a Leap

Claude Opus 4.8 is Anthropic's incremental May 2026 flagship update, positioned by Anthropic as "a modest but tangible improvement" that "builds on Opus 4.7 with improvements across benchmarks." It keeps the same headline pricing as Claude Opus 4.7 ($5 per million input tokens and $25 per million output tokens), adds a faster, higher-cost "fast mode," and ships measurable reliability and honesty gains rather than a step-change in raw capability.

That is the factual frame. What follows is a field report, not a launch recap. I have been running Opus 4.8 inside our daily content operations at ThePlanetTools.ai for a few days, and I want to be honest about what it does well, where it still trips, and what I would tell anyone deciding whether to switch today. Some of this is positive. Some of it is not. The point of a field report is that both can be true at once.

The short version: a correction, not a leap

If you only read one section, read this one. In our reading, Opus 4.8 is a tidy-up release. Anthropic itself frames it as a "modest but tangible improvement," and the benchmark data backs that framing rather than contradicting it. The largest verifiable jump is on the harder coding benchmark, SWE-bench Pro, where the score moves from 64.3% on Opus 4.7 to 69.2% on Opus 4.8 — a roughly 4.9-point gain. On the more familiar SWE-bench Verified, the published table moves only from 87.6% to 88.6%, which is close to the ceiling. On GPQA Diamond, the score actually slips slightly, from 94.2% to 93.6% — a benchmark that is effectively saturated, where a small dip is noise rather than regression.

So the honest summary is this: the raw "is it smarter" needle barely moved. What moved is reliability — fewer code defects slipping through, better self-checking, a more pleasant working tone. For us, that combination matters more than another point on a saturated leaderboard. For other people, with other workflows, it clearly does not. I will get to why.

What actually changed under the hood

Three things stand out in our production use over the past few days.

One: the tone is genuinely better. Opus 4.7 had a working register I can only describe as the slightly annoying intern — eager, a little defensive, prone to over-explaining what it had just done. Opus 4.8 is more conversational and noticeably less grating to work alongside for hours at a stretch. This is subjective, but when you live inside a model all day, ergonomics are not a luxury. They change how much you trust the output without re-reading every line.

Two: the 4.7 reliability bugs are reduced. The two failure modes that hurt us most on Opus 4.7 were premature stopping (the model would declare a long task finished when it was not) and confident over-reporting (it would claim it had analyzed fifty documents when it had actually read fifteen). On Opus 4.8 both happen less. Anthropic's own claim aligns with what we see: it states the model is "around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked." In our hands the self-checking is visibly stronger.

Three — and this is the honest caveat — reduced is not eliminated. Anthropic itself does not claim these behaviors are gone, and neither will I. Their own model documentation acknowledges that occasional early-stopping persists and that the model can still be over-eager about deleting files. We have seen both, just less often than on 4.7. If you read marketing copy as "fixed," you will be disappointed. The accurate word is "reduced."

Benchmarks: the numbers, with the asterisks intact

Benchmarks are useful only when you compare like for like. The single most common mistake in launch coverage this cycle has been opposing scores from different benchmark versions as if they were the same test. They are not. Opus 4.6 and 4.7 published Terminal-Bench 2.0; Opus 4.8 and MiniMax M3 report Terminal-Bench 2.1. Those are different tests. Putting them in the same column would be misleading, so the table below notes the version wherever it differs and uses an em dash ("—") wherever a model has not published a comparable figure. I did not invent a single cell.

Claude line: 4.6 to 4.7 to 4.8

* The 88.6% SWE-bench Verified figure comes from Anthropic's own comparison table as relayed in press coverage. It has not been independently verified by a third party, so treat it as a vendor-reported number, not a settled fact.

Cross-model snapshot (like-for-like only)

Read that table carefully. On SWE-bench Pro — the harder, less-saturated coding test — Opus 4.8 at 69.2% sits clearly ahead of the rest of this field. On GPQA Diamond it is essentially tied with Gemini 3.1 Pro, both bumping the ceiling. The Terminal-Bench rows are deliberately not compared head to head because the versions differ. That is the whole discipline: a benchmark name without its version number is a half-truth.

Our production experience: the long-horizon case

Here is where use-case matters more than any leaderboard. Our work is long-horizon agentic: multi-step content operations, cross-referencing many sources, orchestrating subtasks that run for a long time before they produce anything useful. That is precisely the regime where Opus 4.8's improvements land.

When a task runs for thirty minutes across dozens of files, the cost of a model that quietly stops early or overstates its progress is enormous — you do not catch it until much later, and by then you have built on a false foundation. The reduction in those two specific failure modes is, for us, worth more than any headline benchmark. In our production use, Opus 4.8 finishes long jobs more often than 4.7 did, and lies to itself less about what it actually completed.

I want to be precise about the claim, though: this is our workflow talking. The same traits that help a long-horizon agent can be invisible — or actively annoying — in a quick, turn-by-turn chat. That is not a contradiction. It is the single most important thing to understand about this release, and it explains the split you are about to read.

What would prove me wrong

If I only told you our positive experience, I would be selling you something. So here, in good faith, is the other side — and a community that is genuinely split.

The positive camp looks a lot like us: long-horizon agentic users who value the honesty and self-checking gains, who like the flat pricing carried over from 4.7, and who appreciate the faster 2.5x "fast mode" option when latency matters more than cost. For this group, 4.8 is a clear, if modest, win.

The disappointed camp is real and should not be hand-waved away. On Hacker News, a thread titled "Ask HN: Is Claude Opus 4.8 broken?" collected reports of file-reading failures, hallucinated file paths, repeated tool-use errors, and noticeably slow responses, with at least one commenter describing a session that burned through a large token budget with little to show for it. A separate thread asked whether others were seeing serious degradation in developer experience on 4.8, citing untrustworthy behavior and unintended file deletions. Some users say they reverted to Opus 4.7. Others complain specifically about latency, because the default high-effort setting makes the model slower out of the box. And a vocal subset reports the opposite of our experience: that 4.8 ignores explicit instructions, and that its writing has regressed compared with the older Opus 4.6, which they remember more fondly.

So why does the lived experience diverge so sharply? In our reading, it comes down to use-case, not user error. Long-horizon agentic work — the regime where careful self-checking and finishing the job matter most — is exactly where 4.8 shines. One-shot tasks, writing-heavy prompts, and rapid turn-by-turn iteration are exactly where the slower default effort, the occasional residual file-deletion zeal, and the changed writing voice hurt. If your day is the second kind, the disappointed reports will sound right to you, and you are not wrong.

The hold-over theory (our speculation, scoped)

This next part is speculation, and I am labeling it as such. In our reading, 4.8 has the shape of a hold-over release: a careful consolidation of everything Anthropic can reliably ship right now, while a genuinely larger model is kept back for a future, bigger jump. The evidence is circumstantial — a near-saturated headline benchmark, a "modest but tangible" framing from Anthropic itself, gains concentrated in reliability rather than raw capability. None of that proves a bigger model exists. It is a pattern we have seen before in this industry, and it is how we are reading the tea leaves. Treat it as an informed hunch, nothing more.

Our practical advice to the community

This is the part I actually care about — the working-operator counsel, not the score-sheet.

If your work is routine and reliability is everything, and you can still run Opus 4.6, stay on it. This will be controversial, but in our experience — corroborated by a chunk of the community — Claude Opus 4.6 has ended up more dependable than both 4.7 and 4.8 for ordinary "do this, then do that" instruction-following. With 4.6, you say "do this, do that," and it does it cleanly. If that is your day, the safest move is often no move at all.

If you want to push speed on lower-stakes projects, test 4.8 — but make backups first. Zip your project. Copy your database. This is community advice worth repeating because the residual file-deletion zeal is a genuine, documented behavior, not a rumor. A backup turns a worst-case 4.8 session into an annoyance instead of a disaster. We do this as a matter of routine now, and you should too.

Budget time for recalibration. The way you talk to 4.8 is not the way you talked to 4.6. The model responds to different framing, and the habits you built on the older release need to be re-learned. Do not judge 4.8 on your first hour with it — judge it after you have adjusted your prompting to it.

A prompting insight: stop telling it what not to do

Here is the single most useful prompting habit we have picked up while recalibrating to 4.8, offered as a practical observation rather than a law. Avoid negative instructions. "Do not use bullet points." "Do not think about X." These backfire, for the same reason "do not think about a pink elephant" fails — the negated concept is exactly what gets anchored in attention.

Opus 4.8 feels, to us, closer to human-like cognition than earlier releases, and it responds best to positive framing: say what you want, not what you want avoided. Instead of "do not write a list," try "write this as flowing prose in three paragraphs." This is consistent with Anthropic's own prompting guidance, which generally favors clear, affirmative instructions over prohibitions. It is not a magic switch and it will not fix a bad prompt, but on 4.8 the difference is real enough that we changed our internal prompt templates because of it.

Deep dive: ultracode and the /workflow command

The capability we lean on hardest is Dynamic Workflows inside Claude Code, and it is worth explaining concretely because the marketing language undersells what it actually does day to day.

A Dynamic Workflow is multi-agent orchestration: instead of one model grinding through a task in sequence, the system spawns parallel sub-agents that each take a slice of the work and run at the same time. For data analysis and cross-referenced research — comparing many sources, reconciling conflicting numbers, building a single answer out of a dozen partial ones — this is genuinely exceptional. It is the difference between one researcher reading ten papers in series and ten researchers reading one each and reporting back.

The /workflow command is the part that makes this trustworthy rather than a black box. Run it and you get the detailed report from each individual agent — what it was asked, what it found, what it concluded — plus a global summary of what all the agents did together. For a long-horizon job, that visibility is the whole game: you can see which sub-agent reached a shaky conclusion instead of squinting at one monolithic answer and hoping. "Ultracode" is the maximum-effort mode layered on top — full effort budget plus dynamic orchestration — and it is what we reach for when a job is complex enough to justify the extra tokens. For the cross-analysis work we do, this is where 4.8 earns its keep, separate from any benchmark. If you want the feature-level breakdown of what shipped on launch day rather than this field report, our companion piece on Claude Opus 4.8's dynamic workflows and ultracode in Claude Code goes deeper on the mechanics.

So, should you upgrade?

Honestly, it depends on what you do — which is an unsatisfying answer, so let me make it concrete. If you run long agentic jobs and value finishing-the-task reliability and honesty over raw novelty, 4.8 is a worthwhile, low-risk update at the same price you already pay. If your work is one-shot, writing-heavy, or fast turn-by-turn, and 4.6 is still available to you, you may be happier staying put — or dropping to a lighter model like Claude Sonnet 4.6 for the cheap, fast iterations — and revisiting 4.8 after a recalibration session. Either way: make backups before you let any new model loose on a project you care about. That advice never expires.

Frequently Asked Questions

Is Claude Opus 4.8 a major upgrade over Opus 4.7?

No. In our reading it is an incremental correction, not a leap, and Anthropic itself calls it "a modest but tangible improvement" that "builds on Opus 4.7." The biggest verifiable gain is on SWE-bench Pro (64.3% to 69.2%); SWE-bench Verified barely moves (87.6% to 88.6%, a vendor-reported figure) and GPQA Diamond slightly dips (94.2% to 93.6%) on an already-saturated test. The real wins are reliability and honesty, not raw capability.

What are the known problems with Claude Opus 4.8?

Reduced but not eliminated: Anthropic acknowledges occasional early-stopping and over-eager file deletion still occur. Community reports on Hacker News describe file-reading failures, hallucinated file paths, repeated tool-use errors, slow default responses, and at least one heavily consumed token budget. Some users also report 4.8 ignoring explicit instructions — the opposite of our experience — which we attribute to differences in use-case.

Should I stay on Claude Opus 4.6 instead?

For routine work where reliability is everything, and if you can still run Opus 4.6, yes — stay on it. In our experience, corroborated by part of the community, 4.6 has proven more dependable for plain "do this, then do that" instruction-following than both 4.7 and 4.8. 4.6 does what you tell it cleanly. Switch to 4.8 mainly to push speed on lower-stakes projects, and back up your work first.

How is Claude Opus 4.7 different from Opus 4.8 in daily use?

Opus 4.7 had two failure modes that hurt us most: stopping a long task prematurely, and over-reporting progress (claiming it analyzed fifty documents when it read fifteen). Opus 4.8 reduces both — Anthropic says it is "around four times less likely" to let code flaws pass unremarked — and its conversational tone is more pleasant than 4.7's slightly annoying register. The behaviors are reduced, not gone.

Why do some users love Opus 4.8 while others say it is broken?

Use-case, in our reading, not user error. Long-horizon agentic work — where careful self-checking and finishing the job matter most — is where 4.8 shines, so those users (including us) are positive. One-shot tasks, writing-heavy prompts, and fast turn-by-turn iteration are where the slower default effort, residual file-deletion zeal, and changed writing voice hurt — so those users report regression and slowness. Both experiences are real.

How much does Claude Opus 4.8 cost?

Opus 4.8 keeps the same headline pricing as Opus 4.7: $5 per million input tokens and $25 per million output tokens. Anthropic also offers a higher-cost "fast mode" at $10 per million input tokens and $50 per million output tokens, plus an effort control that lets you choose how hard the model works on a response, with higher settings using more tokens for better results.

What is ultracode in Claude Code?

Ultracode is the maximum-effort mode in Claude Code: it combines the full effort budget with Dynamic Workflows orchestration. It is what we reach for when a job is complex enough to justify the extra tokens — typically cross-referenced research or multi-step data analysis where you want the model to spend more rather than rush to an answer.

What does the /workflow command do?

The /workflow command surfaces the detailed report from each individual sub-agent in a Dynamic Workflow — what each was asked, what it found, what it concluded — plus a global summary of what all the agents did together. It turns multi-agent orchestration from a black box into something you can audit, so you can see which specific sub-agent reached a shaky conclusion instead of squinting at one monolithic answer.

What are Dynamic Workflows in Claude Code?

Dynamic Workflows are multi-agent orchestration: instead of one model working in sequence, the system spawns parallel sub-agents that each take a slice of the task and run at the same time. They are exceptional for data analysis and cross-referenced research — comparing many sources and reconciling them into one answer — which is the long-horizon work where Opus 4.8 earns its keep.

Should I make backups before using Opus 4.8?

Yes — this is community advice worth repeating. Zip your project and copy your database before letting 4.8 loose, because over-eager file deletion is a documented residual behavior, not a rumor. A backup turns a worst-case session into an annoyance instead of a disaster. We treat this as routine now regardless of which model we run.

What is the best way to prompt Claude Opus 4.8?

In our experience, use positive framing: tell the model what you want it to do, not what to avoid. Negative instructions like "do not use bullet points" tend to backfire the same way "do not think about a pink elephant" does — the negated concept gets anchored. This lines up with Anthropic's own prompting guidance favoring affirmative instructions. It is an observation, not a law, but on 4.8 the difference was real enough that we rewrote our prompt templates.

Is there a bigger Claude model coming after Opus 4.8?

We do not know, and nobody outside Anthropic does. Our scoped speculation: 4.8 reads like a hold-over release — a careful consolidation shipped while a larger model is held back for a future, bigger jump. The evidence is circumstantial (a near-saturated headline benchmark, a "modest" framing, gains concentrated in reliability). Treat it as an informed hunch, not a fact.