Claude Sonnet 5: Anthropic's most agentic model, Opus performance at Sonnet price
🔎 Why Sonnet 5 changes the game for AI agents
On June 30, 2026, Anthropic unveiled Claude Sonnet 5. Not a simple refresh, but a paradigm shift in the company's strategy. The stakes go beyond the release of a new model: it is Anthropic's response to an escalating AI price war.
OpenAI was pushing GPT-5.6 Sol as the reference agentic model. Google was aligning Gemini 3.5 Flash in the same segment. Anthropic had to respond, but not just any way. Rather than stacking capabilities into Opus 4.8, the company chose to compress agentic capabilities into a mid-range model.
The result: Sonnet 5 reaches 63.2% in agentic coding, representing 91% of Opus 4.8's score (69.2%), for a cost reduced by more than half after the launch period. A tester quoted by TechCrunch summarizes the situation: where older Sonnets stalled on complex tasks, Sonnet 5 completes them.
The essentials
- Near-Opus performance: 63.2% in agentic coding (vs 69.2% for Opus 4.8), and even slightly surpasses Opus 4.8 in knowledge work according to Anthropic benchmarks.
- Aggressive pricing: $2/MTok input and $10/MTok output until August 31, 2026, then $3/$15 — making it 40% cheaper than Opus 4.8 according to Le Devoir.
- Native agentic: Sonnet 5 is designed to execute multi-step workflows, use tools, and verify its own outputs — the exact profile of what companies are deploying in production.
- Immediately available: default model on Free and Pro plans since June 30, and accessible via API, Claude Code and Cowork.
Recommended tools
| Tool | Main usage | Price (June 2026, check on platform.claude.com) | Ideal for |
|---|---|---|---|
| Claude Sonnet 5 | AI agents, coding, knowledge work | $2/$10 per MTok (promo) then $3/$15 | Developers and product teams |
| Claude Opus 4.8 | High-precision critical tasks | ~$5/$25 per MTok | Cases where the Opus margin is necessary |
| Claude Code | Coding agent in CLI | Included in Pro/Team plans | Individual developers |
Benchmarks: what Sonnet 5 is really worth compared to Opus 4.8
The numbers speak for themselves. Anthropic has released agentic benchmarks that place Sonnet 5 in an unusual zone for a mid-range model: the one where users genuinely hesitate between Sonnet and Opus.
In agentic coding, Sonnet 5 scores 63.2%. That is 5.1 points below Opus 4.8 (69.2%), but 5.1 points above Sonnet 4.6 (58.1%). The gap with the previous model is significant. The gap with Opus is real but may not justify the extra cost for most use cases.
The real surprise comes from knowledge work. On these benchmarks, Sonnet 5 slightly outperforms Opus 4.8. This is a strong signal: for research, synthesis, and professional reasoning tasks, the most expensive model is no longer automatically the best.
Sonnet 5 also beats Haiku 4.5 and Sonnet 4.6 across all published agentic benchmarks. No exceptions. This makes Anthropic's internal hierarchy clearer: Haiku for speed, Sonnet 5 for agentic, Opus 4.8 for the extreme.
Agentic coding comparison
| Model | Agentic coding score | Positioning |
|---|---|---|
| Claude Opus 4.8 | 69.2% | Premium, critical tasks |
| Claude Sonnet 5 | 63.2% | Agentic mid-range |
| Claude Sonnet 4.6 | 58.1% | Former mid-range |
In the broader landscape, Claude Opus 4.7 (Adaptive) scores 94.3 on general LLM benchmarks, and OpenAI's GPT-5.5 dominates at 98.2. But these general scores do not capture agentic specificity. On multi-step execution tasks with tool use, Sonnet 5 holds its own against globally higher-ranked models.
Pricing and strategy: the low-cost agentic war
Sonnet 5's pricing positioning is no minor detail. It is the heart of the strategy. Anthropic is launching at $2/MTok input and $10/MTok output. After August 31, 2026, rates rise to $3/$15. Even at the full rate, it is 40% cheaper than Opus 4.8 (~$5/$25).
For companies running agents in loops — hundreds of thousands of tokens per session, thousands of sessions per day — this difference translates to tens of thousands of dollars per month. Le Devoir points out that this launch comes in a context of a general surge in AI costs. Anthropic is playing the volume card: a cheaper model per token, but designed to consume more of them (multi-step agents).
API Pricing Grid (June 2026)
| Model | Input (per MTok) | Output (per MTok) | Agentic value for money |
|---|---|---|---|
| Claude Sonnet 5 (promo) | $2 | $10 | Excellent |
| Claude Sonnet 5 (after Sept. 2026) | $3 | $15 | Very good |
| Claude Opus 4.8 | ~$5 | ~$25 | Good for critical cases |
| Claude Sonnet 4.6 | ~$3 | ~$15 | Outpaced by Sonnet 5 |
The batch rate for Sonnet 5 is set at $3.75/MTok input starting in September, an interesting point for companies processing batches of documents offline.
Faced with OpenAI GPT-5.6 Sol which is also pushing hard on agentic, and Google's Gemini 3.5 Flash which is targeting the same cost/performance niche, Anthropic could not afford an agentic model sold at Opus pricing. Sonnet 5 is the weapon in this battle.
What makes Sonnet 5 truly "agentic"
The term "agentic" has become a buzzword. But Anthropic defines it precisely: the ability to plan a multi-step task, use external tools, correct its own errors along the way, and produce a verified final result.
Sonnet 5 improves on every dimension compared to Sonnet 4.6. Testers report that it completes tasks where Sonnet 4.6 would stop halfway through. It verifies its own outputs before validating them. It asks for less human guidance between steps.
For developers building agents with Claude Code or via the API, this means longer workflows without intervention. A refactoring agent that, with Sonnet 4.6, would stop after three files can now process eight in one go with Sonnet 5.
The model is also available in Cowork, Anthropic's collaborative interface. This allows teams to delegate knowledge work tasks — research, document synthesis, report preparation — with an unprecedented level of autonomy for a Sonnet model.
If you are looking to compare the best options for this type of workflow, our monthly comparison of the best LLMs details the strengths and weaknesses of each model in real-world scenarios.
Sonnet 5 vs Opus 4.8 : when to choose which
The question everyone is asking: does Sonnet 5 replace Opus 4.8? No. But it drastically reduces the cases where Opus is necessary.
Choose Sonnet 5 for: everyday coding agents, knowledge work workflows, multi-step tasks with tool use, and any case where cost per session matters. It's the default choice for 80% of agentic use cases.
Choose Opus 4.8 for: offensive security tasks (where Sonnet 5 is clearly inferior according to the System Card), cases where every percentage point of accuracy counts, and projects with strict regulatory constraints.
The Claude Opus 4.8 remains Anthropic's most capable model in absolute terms. But Sonnet 5 makes the cost/performance trade-off almost painless for the majority of users.
For developers hesitating between ecosystems, the Claude vs ChatGPT comparison provides an overview of both platforms beyond just the model criterion.
Security and limits: what the System Card reveals
The Claude Sonnet 5 System Card provides a nuanced insight. Anthropic has enabled cyber safeguards by default — less strict than those of the Fable 5 model, but more restrictive than no safeguards at all.
Sonnet 5's overall rate of undesirable behaviors is lower than that of Sonnet 4.6. Good news. But it remains higher than that of Opus 4.8 and Mythos Preview. This confirms Anthropic's security hierarchy: the more powerful the model, the finer the guardrails, but the more rigorous the upstream quality control.
On offensive cyber tasks specifically, Sonnet 5 is clearly below Opus 4.8 and Mythos 5. This is a deliberate choice: Anthropic does not want a mid-range model that can be easily diverted. Cybersecurity companies that used Opus for offensive simulations will not find a replacement in Sonnet 5.
Resistance to prompt injections is improved compared to Sonnet 4.6. Hallucinations and sycophancy (the tendency to agree with the user even when they are wrong) are reduced. Qualitative improvements that are felt in daily use more than they are measured in benchmarks.
New tokenizer: the hidden effect on costs
A technical detail with major financial consequences: Sonnet 5 uses a new tokenizer. The same inputs produce between 1.0x and 1.35x more tokens than with the previous tokenizer.
Concretely, a prompt that counted 1,000 tokens with Sonnet 4.6 can count 1,350 with Sonnet 5. At first glance, this seems negative — more tokens billed for the same content. But Anthropic compensated with a lower price per token.
The final calculation depends on your mix of languages and content types. For French, which is generally more "expensive" in tokens than English, the new tokenizer can cut both ways. Test on your real-world use cases before migrating to production.
For French-speaking users, this tokenizer issue interacts with French language quality. If you are comparing options, our guide to the best LLMs in French can help clear things up.
Competitive positioning: the answer to GPT-5.6 Sol and Gemini 3.5 Flash
The context of this launch is essential. OpenAI GPT-5.6 Sol was positioned as OpenAI's benchmark agentic model. Google responded with Gemini 3.5 Flash, tailored for the same use cases at a low cost.
Anthropic arrived with Opus 4.8 at the top, but at a price that reserved it for premium budgets. Sonnet 4.6 was too weak in agentic to compete. A mid-range soldier was missing. Sonnet 5 is that soldier.
In the overall LLM hierarchy, OpenAI's GPT-5.5 dominates at 98.2, followed by Gemini 3 Pro Deep Think at 95.4 and Claude Opus 4.7 (Adaptive) at 94.3. Sonnet 5 is not aiming for this overall leaderboard. It targets the specific agentic ranking, where the cost/performance ratio matters more than the raw score.
For developers evaluating all options, the landscape of the best LLMs for coding and the best free LLMs remains relevant for comparing access.
Impact for developers and businesses
For individual developers, Sonnet 5 is immediately accessible. It is the default model on Claude's Free and Pro plans. No need to pay more to access it. In Claude Code, it replaces Sonnet 4.6 as the base engine for coding agents.
For businesses, the calculation is different. Sonnet 5 opens up the possibility of deploying agents in production at a sustainable cost. An agent that would cost $100 per day with Opus 4.8 can cost $40-60 with Sonnet 5, for an agentic performance loss of around 8-9%.
Teams building multi-agent systems — for example, a research agent, a code agent, a review agent — can mix Sonnet 5 for standard tasks and Opus 4.8 for critical steps. This hybrid architecture is exactly what Anthropic's pricing makes rational.
Businesses looking to host their own models for privacy reasons can also look towards the meilleurs LLM locaux or follow our guide d'installation d'un LLM en local for on-premise alternatives.
❌ Common mistakes
Mistake 1: Treating Sonnet 5 as a simple upgrade to Sonnet 4.6
This is not an incremental iteration. It is a model with a different agentic architecture. The jump from 58.1% to 63.2% in agentic coding is bigger than it looks, because agentic tasks are non-linear: past a certain threshold, the model starts completing entire workflows that it used to abandon.
Mistake 2: Migrating to production without testing the new tokenizer
The new tokenizer changes the token count of your existing prompts. A direct migration without adjustment can increase your costs by 10 to 35% even before considering the price per token. Test on a representative sample of your actual queries.
Mistake 3: Using Sonnet 5 for offensive security tasks
The System Card is clear: Sonnet 5 is significantly less performant than Opus 4.8 and Mythos 5 on offensive cyber tasks. If this is your use case, stay on Opus.
Mistake 4: Ignoring the end of the promo period
The launch pricing (2$/10$) expires on August 31, 2026. Companies that calibrate their costs to this rate and forget the transition to 3$/15$ risk a nasty budget surprise in September.
❓ Frequently Asked Questions
Does Claude Sonnet 5 replace Claude Sonnet 4.6?
Yes, for all practical purposes. Sonnet 5 beats it on all agentic benchmarks and becomes the default model on Free and Pro. Sonnet 4.6 remains available in API but no longer has any reason to be chosen for new projects.
Is Sonnet 5 free?
The model is accessible for free via Claude's Free and Pro plans, with usage limits. For unlimited API usage, standard pricing applies ($2/$10 per MTok on promo, then $3/$15).
What is the difference between Sonnet 5 and Opus 4.8 in practice?
About 6 percentage points in agentic coding, better security on sensitive tasks for Opus, and a 40% lower price for Sonnet 5. For 80% of use cases, the difference does not justify the Opus premium.
Does the new tokenizer increase my costs?
Not necessarily. Although the same inputs produce more tokens (1.0-1.35x), the lower price per token partially compensates. The net result depends on your type of content and your languages.
Is Sonnet 5 suitable for coding agents in production?
Yes, that is precisely its primary use case. With 63.2% in agentic coding and the ability to complete multi-file tasks, it is designed for this. For the most critical cases, Opus 4.8 remains the most reliable option.
✅ Conclusion
Claude Sonnet 5 does exactly what Anthropic wanted: offer 90% of the agentic power of Opus 4.8 at less than half the price. For developers and businesses, this is the signal that agentic AI is scaling without breaking the bank. If you are still using Sonnet 4.6 in production, migration should be your next priority — and if you are hesitating between models for your agents, our selection of the best LLMs for AI agents will help you choose the right combination.