|
researchaudio.io · Issue 15 · 2026-06-30
|
|
Headline comparison
CyberGym: Sonnet 5 is worse than 4.6.
Launch price: $2 / $10. Standard price starts Sep 1.
The system card and the launch post disagree. The chart agrees with the skeptics.
|
|
Anthropic shipped Claude Sonnet 5. The system card says it is worse at cybersecurity than 4.6.
|
|
The launch pricing is $2 / $10 per million tokens, and ends August 31, 2026. Then the price doubles.
|
|
A Hacker News reader summed up the mood within an hour of the post going live. From @wolttam, currently the top skeptical thread: “I didn't think they'd actually release a model that was worse than the open-weight frontier and at a higher price-point. Wow.”
That is the editorial line. Anthropic's launch post leads with three claims: Sonnet 5 is the “most agentic Sonnet model yet,” “performance is close to that of Opus 4.8, but at lower prices,” and it is a “substantial improvement over its predecessor, Sonnet 4.6” on agentic tasks. The system card Anthropic links from that same post says the opposite on at least one of those three. The card quotes itself: “On CyberGym vulnerability discovery, Claude Sonnet 5 is less capable than Sonnet 4.6, and far less capable than Opus 4.8 and Mythos 5.” Read straight, that is a regression on a security-relevant benchmark the post never mentions by name.
Then there is the chart Anthropic does include, the cost-perf scatter on BrowseComp and OSWorld-Verified. The post says Sonnet 5 “covers a much wider range of cost-performance options than Opus 4.8.” The community read, from @andai: “Opus 4.8 beats Sonnet 5 on the pareto frontier in several of their graphs.” If that reading is right, the headline “close to Opus at lower prices” is partially inverted. Sometimes Opus is the lower-priced option for the same work, because Opus's medium-effort line is already below Sonnet 5's higher-effort line on the same axis.
Both of those findings are inside Anthropic's own material. Neither is a community invention.
|
|
Section
What Anthropic says
|
|
The June 30, 2026 post frames Sonnet 5 around three pillars: most agentic Sonnet yet, close to Opus 4.8 at lower prices, and substantial improvement over 4.6 on agentic tasks.
The availability and pricing footnote is the part the post buries at the bottom: default model for Free and Pro, available to Max, Team, and Enterprise, also in Claude Code and the Platform. Introductory pricing of $2 per million input / $10 per million output through August 31, 2026, then $3 / $15 standard. There is a tokenizer change baked in (Sonnet 5 tokens are 1.0–1.35× the same input on 4.6), so the cost-neutral framing is approximate, not exact.
The safety section is the other half. Anthropic says Sonnet 5 has a lower rate of misaligned behavior than Sonnet 4.6 on its automated behavioral audit, with lower hallucination, lower sycophancy, and better resistance to prompt injection. It also flags that Sonnet 5 is “somewhat higher” in misaligned behavior than Opus 4.8 and Claude Mythos Preview, and that it cannot develop a working exploit on the Firefox 147 benchmark (0.0% full success, slightly higher partial success than 4.6). It is launching with the same real-time cyber safeguards as Opus 4.7 and 4.8.
The system card, linked in the same post, is where the CyberGym quote comes from. The card and the post are not in conflict on safety in general. They are in conflict on the “improvement over Sonnet 4.6” framing for cybersecurity specifically.
|
|
Section
What the numbers actually show
|
| Metric |
Sonnet 4.6 |
Sonnet 5 |
Opus 4.8 |
| OSWorld-Verified (computer use) |
78.5% (revised) |
chart hero, above 4.6 |
high-effort region |
| BrowseComp (agentic search) |
chart baseline |
strictly above on most effort levels |
matches Sonnet 5 at high effort |
| CyberGym vulnerability discovery |
reference |
less capable (system card) |
far more capable |
| Firefox 147 exploit (no safeguards) |
0.0% full |
0.0% full, slightly higher partial |
substantially better |
| HLE (no tools / with tools) |
34.6% / 46.8% (revised) |
not quoted in post |
higher |
|
| Pricing |
$ / M input |
$ / M output |
Notes |
| Sonnet 5 launch |
$2 |
$10 |
through Aug 31, 2026 |
| Sonnet 5 standard |
$3 |
$15 |
after Aug 31, 2026 |
| Opus 4.8 reference |
$5 |
$25 |
higher-end frontier |
| GLM 5.2 (open-weight) |
sub-$1 |
sub-$2 |
744B params, self-host |
|
|
Two things to notice.
First, the cost-perf story is more complicated than the headline “close to Opus at lower prices.” At medium effort, Sonnet 5 undercuts Opus 4.8 on BrowseComp and OSWorld-Verified per Anthropic's own chart, which is the legitimate part of the post. At higher effort levels, the curves cross. @andai's reading is that the higher-effort region of Sonnet 5 is dominated by Opus 4.8 — Opus is both lower-priced and stronger. @doctoboggan says it more bluntly: “if Sonnet 5 medium isn't good enough for you, switch models, not effort levels.”
Second, the CyberGym regression is not a chart interpretation. The system card states it directly. The post's “safer to use in agentic contexts” framing is a generalization that does not survive contact with the cybersecurity evaluation specifically.
Sonnet 5 vs Sonnet 4.6: only eval where 5 is WORSE
CyberGym vulnerability discovery
4.6 [█████████] reference
5 [██████ ] less capable
Opus [████████████] far more capable
Sonnet 5 vs Opus 4.8: the "close to Opus at lower prices" claim
BrowseComp at medium effort (illustrative, from Anthropic chart)
Sonnet 5 [████████ ] 72% (approx) $2 in / $10 out
Opus 4.8 [██████████] ~80% $5 in / $25 out
-> at medium, Sonnet 5 is ~60% the cost, ~90% the score
BrowseComp at high effort
Sonnet 5 [██████████] ~80% $3 in / $15 out (standard)
Opus 4.8 [██████████] ~80% $5 in / $25 out
-> at high effort, Opus 4.8 medium effort is comparable
|
|
|
|
Section
How it actually works
|
|
In plain English: you tell Claude Sonnet 5 what you want, it picks a plan, opens a browser or terminal, tries things, checks whether they worked, and keeps going until it either finishes or decides it cannot — all in one session, on its own, for as long as you let it.
|
|
The deeper version, in three pieces.
Effort levels. The same model runs at four settings: low, medium, high, xhigh. The setting changes how much compute the model spends per turn, how thoroughly it plans, and how often it re-checks its own output. The community reaction is that the new xhigh is closer to Sonnet 4.6 at high than to anything genuinely new. From @alvis: “Today sonnet 5's med level effort is equivalent to sonnet 4.6 low level effort.” Whether that is a bug or a feature depends on what you were paying for in 4.6.
Tool use and planning. Sonnet 5 is built to be the agentic one. It can chain tool calls, write its own test cases, and produce a verified result before reporting back. Several launch partner quotes (Cursor, ClickHouse, Lovable, Eve, Pace) describe the same pattern: a single Sonnet 5 session finishes a multi-step task that older Sonnets abandoned halfway.
The tokenizer change. Sonnet 5 uses a new tokenizer, the same one that shipped with Opus 4.7. The same input maps to more tokens — 1.0× to 1.35×, depending on the content. Anthropic says the introductory price is set so the transition is “roughly cost-neutral.” “Roughly” is the load-bearing word.
The honest caveats are inside the post itself. The chart on BrowseComp was originally based on a simpler methodology; Anthropic updated the post on June 30 to use the standard 10M-token budget with compaction. The current chart is the more favorable of the two. The earlier chart was buried in a “Changelog” footnote.
|
|
|
|
Section
Where it works, where it collapses
|
|
Where it works
Strongest Sonnet yet on agentic coding and multi-step tool use at medium effort, per the chart and the launch partner quotes.
The 1.0–1.35× tokenizer hit is offset by the launch pricing through August 31.
Default model on Free and Pro, so anyone with a Claude account gets the upgrade on login.
Wider effort-level spread (low / medium / high / xhigh) gives more room to tune cost vs quality than 4.6 had.
|
|
Where it collapses
System card explicitly says Sonnet 5 is worse than 4.6 on CyberGym. The “substantial improvement” framing does not survive on that benchmark.
On Anthropic's own charts, higher-effort Sonnet 5 is dominated by medium-effort Opus 4.8 in some regions. The “close to Opus at lower prices” line is only true at medium effort.
Launch pricing is a 60-day window. After August 31, the price doubles to $3 / $15, into a crowded field that already includes GLM 5.2 at 744B parameters.
@microtonal: “the more models are optimized for fully agentic development, the worse they get at assisted development.”
|
|
|
Section
Community reaction
|
|
Hacker News story ID 48736605, posted by @marinesebastian, June 30, 2026. 1,210 points, 743 comments. The thread is the only editorially honest place to read the launch, because Anthropic's own chart is ambiguous and the post buries the contradictions in footnotes.
|
|
“I didn't think they'd actually release a model that was worse than the open-weight frontier and at a higher price-point. Wow.”
@wolttam, HN 48736605, 2026-06-30 — the strongest skeptic in the thread
|
|
|
Skeptic
“I didn't think they'd actually release a model that was worse than the open-weight frontier and at a higher price-point. Wow.”
@wolttam, HN, 2026-06-30
Technical observation
“Opus 4.8 beats Sonnet 5 on the pareto frontier in several of their graphs (Agentic Search, Agentic Computer Use). In other words, for certain tasks, Opus 4.8 is cheaper than Sonnet 5, and does better than Sonnet 5. ... tldr: if you're doing something hard, just use a bigger model.”
@andai, HN, 2026-06-30
System card question
“Wow, seems worse even on price/performance than GLM 5.2, which is only 744b parameters. From the system card: 'On CyberGym vulnerability discovery, Claude Sonnet 5 is less capable than Sonnet 4.6, and far less capable than Opus 4.8 and Mythos 5'”
@conradkay, HN, 2026-06-30 (light formatting, quote excerpted from a longer post)
Access and distribution
“What I starting to hate is that each model's effort level can mean completely different power. Today sonnet 5's med level effort is equivalent to sonnet 4.6 low level effort :/”
@alvis, HN, 2026-06-30
Nuance
“This is much more interesting of a model at $2/$10 (their launch pricing) than at full price. ... I don't think it will make sense at full price.”
@mchusma, HN, 2026-06-30 (lightly edited for length)
Optimist
“Seems to be another great incremental update to the workhorse, nice! I've been using Sonnet instead of Opus for almost all coding tasks for a while now. A little elbow grease to break down tasks and you can spend a lot less money for just about the same output quality.”
@phillipcarter, HN, 2026-06-30
The thread skews skeptical because the contradictions are inside the source material, not invented by the commenters. The strongest single line is @wolttam's, because it captures both the regression (worse than the open-weight frontier) and the cost story (higher price) in one sentence.
|
|
Section
What this means for
|
|
Junior engineer
The real lesson is the effort-level system, not the model. low / medium / high / xhigh is now a per-API-call parameter, and the cost difference between medium and xhigh is more than 4×. Learn to default to medium, escalate per task, and measure your own cost-per-task before assuming the new model is the right tool. Sonnet 5 at medium effort is the more interesting upgrade than Sonnet 5 at xhigh.
|
|
|
Senior engineer
The system card regression on CyberGym is the part to read carefully, not the post. Anthropic shipping a model that is worse than 4.6 on a specific named benchmark is a real signal about how safety evaluations are being weighted relative to capability benchmarks. The “safer to use in agentic contexts” line is a generalization that is true on most axes and not on this one. Build your safety model around the chart, not the headline.
|
|
|
Hiring manager
Sonnet 5 changes the workload shape more than the skill shape. The work that benefits is the kind an agent runs for 30+ minutes: multi-file refactors, account-tier migration, insurance FNOL, legal research. The work that does not benefit is the kind a senior engineer does in 10 minutes with a tight context window. Hiring for “agentic coders” is fine. Hiring for “Sonnet 5 whisperers” is not a job yet.
|
|
|
Founder
August 31 is the actual decision date. The 60-day launch pricing makes Sonnet 5 a reasonable default for new agent products shipping in Q3 2026. After that, the cost story collapses into a crowded open-weight field. If your product's unit economics require Sonnet 5 to stay at $2 / $10, your pricing model does not survive September 1. If your product's unit economics survive $3 / $15, the open-weight option (GLM 5.2, the next Qwen) is one model release away from being the lower-priced option anyway.
|
|
|
Section
The metric that actually matters
|
Sonnet 5 cost (per 1M tokens, 30% input / 70% output blended)
Launch (through Aug 31): $7.60
Standard (after Aug 31): $11.40
Opus 4.8 reference: $20.00
What you actually pay for "Opus 4.8 medium" vs "Sonnet 5 high"
Opus 4.8 medium, 1M blended: $20.00
Sonnet 5 high, 1M blended: $11.40 <-- if this is on the chart
(Sonnet 5 cost rises 50% post-launch; Opus 4.8 cost is unchanged)
The August 31 expiry is the only number that matters
Days until launch price ends: 62
Days until the next Opus: ~30 (@mesmertech estimate)
Days until the next GLM/Qwen: ~45-90
The cost-perf is the story. The August 31 expiry is the deadline. The next Opus is probably the more important release.
|
|
If you share one number from this issue, share this one
“On CyberGym, Sonnet 5 is worse than 4.6. The launch price is $2/$10. It ends Aug 31.”
Anthropic's post says “substantial improvement over 4.6.” Anthropic's system card says the opposite on cybersecurity. Both come from Anthropic.
researchaudio.io · issue 15 · June 30 2026
|
|
Sonnet 5 is a real upgrade on agentic coding, a meh upgrade on the rest, and a $2 / $10 launch that ends in 60 days. The system card and the launch post disagree about cybersecurity. The cost story is only good for two months. The open-weight frontier is the actual competitor, and the frontier is not impressed.
|
|
Section
Reader challenge
|
|
Three questions to sit with this week.
1. If medium-effort Sonnet 5 is dominated by medium-effort Opus 4.8 on BrowseComp at the same price (per @andai), what does the “close to Opus at lower prices” pitch even mean?
2. If the system card says Sonnet 5 is worse than 4.6 on CyberGym, and the post says it is safer in agentic contexts, which document do you trust for a security-sensitive deployment?
3. If GLM 5.2 (744B, open-weight) is comparable on price-perf to a model you rent, what is the API actually selling you?
Next issue: what the August 31 expiry does to agentic-coding product unit economics, and which open-weight release the pricing curve is actually racing against.
|
|
