Just shipped: Anthropic Claude Sonnet 5. The company calls it the “most agentic Sonnet ever”. The model plans, runs terminals and browsers autonomously, and can run long-term tasks.
Today, Sonnet 5 has become the standard model in both Free and Pro plans. Max, Team or Enterprise users may select it. The feature is live on Claude Code, as well.
TL;DR
- Anthropic Sonnet 5 – the mid-tier agentic modelThe gap between Opus 4.8 and 4.9 is now much smaller.
- Sonnet 4.6 is beaten on all published benchmarks: 63.2% SWE-bench Pro, 81.2% OSWorld-Verified, 57.4% HLE.
- Running cheaperPricing for MTok is $2/$10 until Aug 31. After that, it’s $3/$15. Opus 4.8 costs $5/$25.
- The best possible value for low/medium effortsAt xhigh, it is possible to pay more for Opus 4.8 with similar quality.
- Safety is higher than the 4.6 version, but with a deliberately lower cyber capability — Opus stays the pick for accuracy-critical work.
Claude Sonnet 5
The middle product in Anthropic’s line-up is Sonnet. The Opus 4.8 is the top model, while the Haiku 4.5 is the lower one.
Sonnet 5 was launched in February of 2026. The focus of this release, according to Anthropic, is agentic reliability and not a single headline benchmark.
It means, in reality, longer task chain without losing context. This means that a failed tool call can be corrected more quickly. The behavior will be more stable over the course of extended sessions within Claude Code or Cowork.
Models show effort levels as low, medium high and extra high. A higher effort requires more reasoning tokens. Both quality and costs will increase.
Note that Opus 4.7 introduced a new tokenizer for Sonnet 5, which is the one used in Sonnet 5. A text may map roughly to 1.0-1.35 times as many tokens.
Interactive Explainer
Claude Sonnet 5 — Cost & Capability Explorer
Compare published benchmarks and estimate the cost per task across all models. Anthropic launched on June 30, 2026.
Per-task cost estimator
$0.00
per task • $0.00/day • $0.00/mo
Sonnet 5 has an updated tokenizer, the same as Opus 4. The same text can map to roughly 1.0–1.35× more tokens, so the factor is applied to Sonnet 5 only.
Benchmark comparisons
Sonnet 4.6
Sonnet 5
Opus 4.8
Sonnet 5 has a score of 1,618 compared to Opus 4.8’s 1,615. The benchmark used a completely different scale. It is therefore shown as a line rather than an object.
Benchmark
The Anthropic Team published a table of benchmarks comparing Opus 4, Opus 5, and Sonnet 5. Sonnet 5 is superior to its predecessors in all categories. It is a significant improvement over Opus 4.
Sonnet 5 scored 63.2% on agentic coding using SWE-bench pro. Sonnet 4.6 was rated at 58.1%. Opus 4.8 is still the leader at 69.2%.
Sonnet 5 has a score of 81.2% compared to Sonnet 4.6’s 78.5%. It reaches 80.4% on Terminal-Bench 2.0 compared to 67.0%.
Sonnet 5, on Humanity’s Last Exam, scored 57.4%. It’s almost identical to Opus 4, which is 57.9%.
Sonnet 5 has a slight advantage in one area. Sonnet 5 scores 1 618 points on GDPval-AA-v2 knowledge work benchmark compared with Opus 4.8’s score of 1,615.
What is the real trade-off?
Developers should focus on the story of cost-performance. Sonnet 5 represents a significant improvement in comparison to Sonnet 4.6 at all effort levels. Low and medium efforts show the most value.
Sonnet 5 offers quality at a price that is comparable to earlier Sonnets. Opus 4.8 is the leader in accuracy at the very top.
This is the basis for a practical routing policy. Sonnet 5 is the best place to send most of your agentic coding and tool usage, as well as knowledge-based work. Opus 4.8 is reserved for tasks that require accuracy. Haiku 4.5 is best for calls that are high in volume and latency sensitive.
Sonnet 5, Use Cases
Early access partners have described workflows in detail. They map their reports to engineering tasks.
- Software engineering in multiple stepsA tester requested that Sonnet 5 investigate a problem. This was done by writing a reproducing bug test and implementing the fix. This was done in one pass.
- Brownfield debuggingThe model was run by another partner on hard requests. This model tracked down the causes of failures. It sent durable solutions rather than just symptom patching.
- Business automationZapier gave it two jobs. The update was made to the Salesforce account levels, and then an email with a link for enterprise contacts was sent. The entire task was accomplished.
- Agents that use computersPace manages insurance workflows such as submission intake and losses runs. The agents use the same operational systems that teams are already using.
- Data explorationClickHouse agents can query data in real time and generate insights instantly. Analysts can get insights faster with quicker reasoning.
Comparison Table
| Metric/Spec | Sonnet 4.6 | Sonnet 5 | Opus 4.8 |
|---|---|---|---|
| SWE-bench pro: Agentic Coding | 58.1% | 63.2% | 69.2% |
| Terminal-Bench 2.1 | 67.0% | 80.4% | Not reported |
| Computer use (OSWorld-Verified) | 78.5% | 81.2% | Not reported |
| Humanity’s last exam (with tools). | 46.8% | 57.4% | 57.9% |
| Knowledge work (GDPval-AA v2) | Not reported | 1,618 | 1,615 |
| The input price is expressed as a dollar per MTok. | 3 | After 2 introductions, 3 | 5 |
| The output price is expressed as a dollar per MTok | 15 | After 10 minutes, 15 | 25 |
Sont 5 will continue to offer its introductory price until August 31st 2026. After that date, the standard pricing will be $3/$15. The standard prompt cache (cache is read at 0.1x input), as well as the Batch API 50% discount, also applies. Sonnet 5 costs less per token than Gemini 3 Pro and GPT-5.5, but more than Gemini 5 Flash. Anthropic has listed a context window with 1M tokens in the launch of Sonnet 5 on its website. The context numbers for other models are not published here.
Coding Example: Calling Sonnet 5
You can call any Anthropic Model. The model string is changed to claude-sonnet-5.
Import anthropic
Anthropic client is the same as anthropic.() # Reads ANTHROPIC_API_KEY
message = client.messages.create(
model="claude-sonnet-5",
max_tokens=1024,
messages=[
{"role": "user", "content": "Find the race condition in worker.py and ship a tested fix."}
],
)
print(message.content[0].text)Strengths & Weaknesses
Strengths:
- In every category tested, the Anthropic version of Sonnet 4.6 improves.
- Near-Opus quality 4.8 on multiple evaluations at lower prices per token
- On the GDPval AA v2 Knowledge Work Benchmark, Opus 4 8 edges out on its own.
- Sonnet 4.6 has lower rates of hallucination and sycophancy as well as undesirable behavior.
- You only need to swap out the model string.
Weaknesses:
- Even on the most difficult accuracy-critical tasks, Opus 4.8 wins.
- Costs can be higher than Opus 4.8 for xhigh efforts
- New tokenizers can increase token count by as much as 1.35 times
- Opus is sanctioned for cyber work; Cyber capabilities are intentionally kept low
- After August 31st 2026, the standard pricing will be $3/$15.
Claude Sonnet 5 — Community Reaction
Hacker News and X will launch on June 30th, 2026.
Eight reactions and their respective emotions
Positive · 38%
Neutral / mixed · 38%
Negative · 25%
The price/value ratio has received mixed reviews: some praise, while others are unsure about the pricing. These were labeled manually from public posts; Reddit’s two links below are active threads and not counted.
X@ClaudeDevs (official)Positive
“Top-tier performance on coding and tool use at Sonnet pricing” — with a 1M context window.
Hacker NewsPhilipcarterPositive
“Another great incremental update to the workhorse.” Most coding is done using Sonnet instead of Opus.
Hacker NewsMchusmaMixed
The launch price of $2/$10 is far more attractive than the full-priced standard pricing.
X@kimmonismusPositive
“Near Opus 4.8-level performance, but cheaper.” Gains in logic and the use of tools.
Hacker NewsTheaiThe Critical Language
“If you’re doing something hard, just use a bigger model.” Opus gains parts of the Frontier.
Hacker NewsConradkayThe Critical Language
“Seems worse even on price/performance than GLM 5.2” The 744B parameter is a set of parameters.
Hacker Newsmag7269Neutral
“When can we get a new Haiku?” The 4.5 version is almost a year-old and shows its age.
Hacker NewsThe bredrenMixed
When compared with Opus 4, the difference is less pronounced at higher effort levels.
Redditr/ClaudeAI
Launch-day discussion — benchmarks, pricing, and Claude Code impressions from the community.
Redditr/LocalLLaMA
Compare the GLM-5.2 versus K2.7 weights to compare their performance and price.
Since a canonical launch thread was not yet complete at the time of publication, Reddit Cards link to subreddits that were live on launch day. Hacker News cards and X Cards quote public posts that can be linked. The sentiment labels were manually read by an editor, and not automatically scored.
Take a look at the Technical details. Also, feel free to follow us on Twitter Join our Facebook group! 150k+ML SubReddit Subscribe now our Newsletter. Wait! Are you using Telegram? now you can join us on telegram as well.
You can partner with us to promote your GitHub Repository OR Hugging Page OR New Product Launch OR Webinar, etc.? Connect with us
