July 2026 packed four major model launches into ten days: Claude Sonnet 5, GPT-5.6 Sol, GPT-5.6 Luna, and Grok 4.5. Each one changed the cost-to-capability math for anyone building with AI. If you pick models by reading benchmark tables, this month rewrote your decision tree.
Tarsk now recommends twelve models. Four of them matter most right now. Claude Sonnet 5 is the new agentic coding default. GPT-5.6 Sol is the frontier reasoning leader. GPT-5.6 Luna is a budget workhorse that outperforms last month's flagship at one-sixteenth the cost. Grok 4.5 tops every model on the market for agentic tool use.
Claude Sonnet 5: Near-Opus Intelligence, Sonnet Prices
Anthropic released Claude Sonnet 5 on June 30. It replaces Sonnet 4.6 as our recommended mid-tier model.
Sonnet 5 lands close to Opus 4.8 on agentic evals at a fraction of the price: $3 per million input tokens and $15 per million output, compared to Opus 4.8 at $5/$25. SWE-bench Pro jumped from 58.1% (Sonnet 4.6) to 63.2%. Terminal-Bench 2.1 vaulted from 67% to 80.4%. You get strong coding and agent performance without paying Opus prices.
The tradeoff: Anthropic removed manual extended thinking and custom sampling parameters. The new tokenizer produces about 1.3x more tokens for the same text, so your per-task cost runs higher than the list price alone suggests. On GDPval-AA v2, Sonnet 5 scores 1,618 against Opus 4.8's 1,615. They are neck and neck on knowledge work. If you were hoping for a clear Opus-beater at a discount, this is not it. If you wanted Opus-quality agents at 40% less, it delivers.
Customer-facing agents, production coding workflows, long-context document analysis: that is Sonnet 5's territory. The 1M token context window is standard, and adaptive thinking handles simple and complex queries without you needing to configure reasoning effort.
Introductory pricing runs through August 31: $2/$10 instead of
$3/$15. After that, standard pricing kicks in. Use claude-sonnet-5 via
the Anthropic provider.
GPT-5.6 Sol: The Frontier Reasoning Leader
OpenAI shipped GPT-5.6 on July 9 with three tiers: Sol (flagship), Luna (budget), and Terra (the direct GPT-5.5 replacement). We did not add Terra because Sol and Luna bracket it on both capability and cost.
Sol leads the frontier on reasoning. It posts 58.9 on the Artificial Analysis Intelligence Index, behind only Claude Fable 5 at 59.9. It ranks #1 on the Coding Agent Index at 80.0. SWE-bench Verified: 88.1%. Terminal-Bench 2.1: 88.8%, or 91.9% with Ultra Mode running four agents in parallel. If you need maximum intelligence and money is not your first constraint, Sol is the pick.
Pricing: $5 / million input and $30 / million output. Same list price as GPT-5.5, for a model that scores 254 Elo higher on GDPval-AA v2 (1,748 vs 1,494). Cache write pricing, batch processing at half cost, and 90% discount on cache reads are new with this release.
Sol handles hardest reasoning problems, complex multi-step agent pipelines, and research workflows where a wrong answer is expensive. It manages long-context reasoning above 256K tokens at 91.5% MRCR, an area where cheaper models degrade fast.
GPT-5.6 Luna: Flagship-Last-Month Intelligence at Budget Prices
Luna is the quiet star of the July releases. At $1 per million input and $6 per million output, it outperforms GPT-5.5 on nearly every benchmark.
Luna matches or beats GPT-5.5, April 2026's flagship, at one-sixteenth the effective cost after caching and batch pricing. SWE-bench Pro: 62.7% vs GPT-5.5's 58.6%. Terminal-Bench 2.1: 84.7% vs 83.4%. It is the fastest GPT-5.6 model at 255 output tokens per second.
The tradeoff: long-context reliability drops above 256K. Luna's MRCR score at 256K-512K is 41.3%. Sol scores 91.5% on the same test. The 1.05M context window exists, but do not depend on it for reasoning across the full span.
High-volume classification, routing, extraction, summarization. Lightweight agents and sub-agents where latency matters more than peak intelligence. Customer support triage. A workload that costs $4,000 a month on Sol costs about $800 on Luna. Product teams notice that math.
Grok 4.5: Best-in-Class Tool Use at Half the Price
xAI launched Grok 4.5 on July 8 under the SpaceXAI banner. It is the first Grok model on Tarsk's recommended list.
Grok 4.5 ranks #1 on τ³-Banking, the leading agentic tool-use benchmark, at 33% against GPT-5.5's 31%. On SWE-bench Pro, it uses about 16,000 output tokens per task. Opus 4.8 uses 67,000. That is 4.2x fewer tokens. At $2/$6 per million tokens, Grok 4.5 costs about 60% less than Opus 4.8 and matches or exceeds it on agentic coding. Intelligence Index: 53.8, near GPT-5.5 territory.
What you give up: the context window is 500K tokens, down from Grok 4.3's 1M. xAI did not publish a full benchmark table with MMLU or HumanEval. No safety card shipped at launch. Grok 4.5 is not available in the EU yet. If your procurement process requires published safety documentation or standard reasoning benchmarks, you run your own evals.
Use Grok 4.5 for agentic coding where tool-use reliability matters above all else. Automated ticket resolution. Cursor or Grok Build users. Anyone who wants Opus-class agentic performance at well under Opus pricing.
The Full Recommended List
All twelve models Tarsk recommends as of July 2026:
- Claude Haiku 4.5 — cheap, fast Anthropic access
- Claude Sonnet 5 — balanced agentic coding
- Claude Opus 4.8 — maximum Anthropic capability
- Claude Fable 5 — Anthropic's Mythos-class reasoning
- GPT 5.6 Sol — OpenAI's frontier flagship
- GPT 5.6 Luna — budget OpenAI with last-gen-flagship smarts
- Qwen3.7 Max — open-weight reasoning
- Gemini 3.5 Flash — Google's efficiency leader
- Kimi K2.6 — long-context specialist
- Deepseek V4 Pro — open-source frontier
- GLM-5.2 — open-weights MIT license, coding contender
- Grok 4.5 — best-in-class agentic tool use
How to Pick
Try these models in Tarsk
Every model on this list is available today. Head to Settings, connect your API key, and pick the model that matches your workload.