Claude Sonnet 5 Just Won the Mid-Tier War. GPT-5.6, Gemini 3 Pro, DeepSeek V4, and Qwen 3.7 Max Are All Chasing It.

Hey guys, Mr. Technology here.

Every other Sonnet launch has been "Sonnet is good but Opus is better." That stopped on June 30, 2026. Claude Sonnet 5 is the first Anthropic mid-tier model that is not playing defense. On Terminal-Bench 2.1 it beats Opus 4.8 outright (80.4% vs 74.6%). On the headline knowledge-work eval it edges Opus on the same harness (1,618 vs 1,615). The pricing is unchanged from Sonnet 4.6 ($3/$15 standard, $2/$10 through August 31). The rest of the mid-tier — GPT-5.6, Gemini 3 Pro, DeepSeek V4 Pro, and Qwen 3.7 Max — is now in chase position. Most of them are not catching up.

If you build LLM-powered software, this is the only model launch you actually need to understand this week. Not because the benchmarks are unprecedented. Because the Sonnet tier is now the de facto frontier for production agentic work, and the rest of the field is being measured against a price-to-performance bar it did not expect to be defending.

What Anthropic Shipped

• Model: claude-sonnet-5, drop-in for claude-sonnet-4-6
• Launched: June 30, 2026
• Context: 1M input, 128K output
• Pricing: $2/$10 introductory through Aug 31, 2026; $3/$15 standard
• Modality: text + vision
• Reasoning: adaptive thinking on by default; manual thinking blocks now return HTTP 400
• Sampling: temperature, top_p, top_k at non-defaults also return HTTP 400
• Tokenizer: new tokenizer; same English text produces ~30% more tokens than 4.6
• Removed: Priority Tier service tier

Three of those bullets are breaking changes for existing production traffic. We will get to those in the next post. This one is about the competitive picture.

The Numbers, Against the Field

Here is Sonnet 5 measured against the rest of the mid-tier in 2026. Numbers are from each vendor's published system cards and release notes; comparisons use the same harness where one is public.

Read that table before you argue with me about open weights.

DeepSeek V4 Pro wins SWE-bench Pro at 64.8%, by 1.6 points. It is the best coding model in this tier on raw benchmark. That is the row DeepSeek's marketing team is going to lead with. It is also the row that misleads the most.

On Terminal-Bench 2.1 — the benchmark that actually captures multi-step agentic work in a real shell — Sonnet 5 leads the entire table at 80.4%. Opus 4.8 is 5.8 points behind. GPT-5.6 Terra, Gemini 3 Pro, DeepSeek V4 Pro, and Qwen 3.7 Max are all behind both Anthropic models. The Sonnet team shipped a model that beats its own flagship on the harness that matters most for agents. The system card does not headline this. The data is in the table.

OSWorld-Verified, Sonnet 5 is 0.8 points behind Opus 4.8, ahead of everything else. Computer use is now a solved problem at the mid-tier if you are inside Anthropic, and clear second place if you are not.

Humanity's Last Exam with tools: Sonnet 5 is essentially tied with Opus 4.8 (57.4% vs 57.9%) and ahead of every other vendor. The reasoning gap that justified the Opus premium is gone for most workloads.

Pricing Reality Check

DeepSeek V4 Pro is $0.55/$2.20. Qwen 3.7 Max is $0.40/$1.60. On pure dollar-per-token, Sonnet 5 loses badly. That is the open-weight pitch.

Here is what that pitch hides:

• DeepSeek V4 Pro's 1M context is paper-only at production-grade latency. Real 1M-context V4 Pro traffic routes through 200K context slots; cache misses crater throughput.
• Qwen 3.7 Max's 256K context is a quarter of what Sonnet 5 ships. "Load a repo, summarize, ask follow-ups" does not fit.
• Token economics on agents are dominated by tool-call output, not input. Sonnet 5 finishes tasks in fewer steps than 4.6 did. DeepSeek's per-token advantage evaporates once your agent does more than five tool calls.
• You pay per failed task, not per token. If V4 Pro takes three attempts to do what Sonnet 5 does in one, the $0.55 input price is irrelevant.

The "$3/$15 is too expensive when DeepSeek is $0.55" argument was already weak in 2025. After Sonnet 5, it is dead.

Where Sonnet 5 Loses

I am not here to sell you a model. Three real weaknesses:

1. Vision is not best-in-class. Gemini 3 Pro still owns multimodal reasoning on documents, charts, and UI screenshots. Sonnet 5's vision is fine; it is not what you pick Gemini for. 2. Tokenizer inflation is a real budget hit for workloads tuned against Sonnet 4.6 token counts. Code completion, autocomplete, anything that fans out into thousands of small completions — expect a 20-30% bill jump on day one of migration, even at the introductory price. 3. Priority Tier is gone. If you were paying the 4.6 priority premium for batch-sensitive workloads (real-time voice, customer-facing chat with latency SLOs), you cannot reach for Sonnet 5. You stay on Opus 4.8 priority or you eat the latency.

The first is a deliberate Anthropic bet that vision is not where agentic value sits. The second is a hidden tax you will discover when the bill arrives. The third is the single feature gap that keeps Opus 4.8 on the menu.

What You Should Actually Do

• On Sonnet 4.6: migrate to Sonnet 5 today. Same price. Materially better on every agentic benchmark. Read the breaking-changes list before flipping the model string in production.
• On Opus 4.8: pull Sonnet 5 into your routing layer and A/B on real traffic for two weeks. It will handle 80%+ of what you send to Opus today at one-fifth the cost. Reserve Opus for the long tail.
• On GPT-5.5 / GPT-5.6 Terra: stay on GPT-5.5 until 5.6 Terra opens up beyond the limited-preview partners. When it ships broadly, it is the only direct competitor. Until then, Sonnet 5 is the default.
• On Gemini 3 Pro: keep it for vision-heavy and Workspace-integrated workloads. For pure agentic coding and tool use, switch to Sonnet 5. The price delta is real; the capability delta is larger.
• On DeepSeek V4 Pro / Qwen 3.7 Max for cost reasons: re-test your actual workload against Sonnet 5 at the introductory price. The token savings are real on short prompts. They almost always disappear once you count retries and human cleanup.

The Bigger Picture

Three things happened on June 30 that nobody is connecting yet.

First, the Anthropic frontier narrowed. Sonnet 5 is essentially Opus 4.8 in capability for most workloads, and beats Opus on Terminal-Bench 2.1. The "Opus is for hard things, Sonnet is for cheap things" mental model from 2025 is over.

Second, OpenAI's GPT-5.6 launched on June 26 but only ~20 vetted partners can touch it. The lab with the deepest distribution is shipping behind a closed door. Anthropic made the opposite call with Sonnet 5 — Free, Pro, Max, Team, Enterprise, API, Claude Code, Bedrock, Vertex. Everyone has it on day one. The competitive moat is not the model anymore; it is the time-to-access.

Third, the open-weight gap on agentic benchmarks widened in the wrong direction. DeepSeek V4 Pro is the only open-weight model in this tier that competes on coding. On everything else — terminal work, computer use, agentic search, knowledge work — open weights are now two tiers behind on agentic evals, not one. The "open weights are catching up" narrative was true in 2025. It is not true in July 2026.

The mid-tier war is not over. GPT-5.6 Terra will land broadly in the next few weeks and it will be a real fight. Gemini 3 Pro will get a reasoning upgrade by August. DeepSeek will ship V4.5 with terminal-tool-use focus. Qwen will keep undercutting on price.

But right now, on July 8, 2026, Sonnet 5 is the model you default to if you are building anything that has to do real work in a real shell against a real codebase at production volume. The bar moved. Most of the field is still running toward where it was.