Summer 2026 frontier releases: tiers, effort settings and a falling price floor

July 2026 produced three releases that frame how companies will buy AI models for the rest of the year: OpenAI's GPT-5.6 family, Anthropic's Claude Opus 5 and Moonshot AI's Kimi K3. The headline benchmark claims are, as usual, mostly vendor-reported, and the OpenAI figures below come from trade-press coverage rather than OpenAI's own pages, which we could not open; the Anthropic figures come from Anthropic's announcement. The more durable change is structural: each release is sold as a menu of price and effort options, not a single model.

OpenAI: one family, three price points

OpenAI released GPT-5.6 on July 9 as Sol (the flagship), Terra and Luna. According to AI Weekly, the per-million-token prices for input and output at launch were $5/$30 for Sol, $2.50/$15 for Terra and $1/$6 for Luna ^[1]. The same report says Terra is being pitched as similar in performance to GPT-5.5 at roughly half the price ^[1]. It also notes a limited preview with roughly 20 government-approved partner organizations before the Commerce Department's Center for AI Standards and Innovation ran additional tests and cleared the broad rollout ^[1].

Let's Data Science reports that on July 30 OpenAI lowered Luna by 80% (to $0.20 input and $1.20 output) and Terra by 20% (to $2 and $12), leaving Sol unchanged ^[2]. OpenAI credited "improvements across model serving and production software" and said Sol-assisted kernel work cut end-to-end serving costs by 20%; the outlet notes that the 20% figure is company-reported and not independently validated ^[2].

Anthropic: price held, argument shifted to cost per task

Claude Opus 5 arrived on July 24 at $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8 ^[3]. Anthropic's case is built on cost-adjusted comparisons. It says Opus 5 at max effort performs within 0.5% of Fable 5's peak score on CursorBench 3.2 at half the cost per task, and surpasses Fable 5's best OSWorld 2.0 result at just over a third of the cost ^[3]. Fable 5 is listed at $10 input and $50 output per million tokens ^[4]. Most of the benchmark results are published as charts without exact scores, so buyers should treat them as directional until they have run their own tests ^[3].

Two details matter operationally. First, customers can tune effort per request; the announcement names "high," "xhigh" and "max" levels and mentions a lowest-effort setting, trading intelligence against tokens ^[3]. Second, Anthropic says it expects its cyber classifiers to intervene around 85% less often than they do for Fable 5, and describes a Cyber Verification Program for security work that safeguards would otherwise impede ^[3].

Moonshot: a large open-weight entrant

Moonshot released Kimi K3 on July 16. It is described as a 2.8-trillion-parameter sparse mixture-of-experts model that activates 16 of 896 experts per token, with a one-million-token context window ^[5]. Coverage differs on active parameters per token (about 50 billion in one report, 104 billion in another), which is a reason to wait for the technical report ^[6][7]. On GDPval-AA v2, Moonshot reports K3 at 1,687, behind Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8) and ahead of Claude Opus 4.8 (1,600) ^[5]. Those are the company's own claims, and early coverage cautioned that parameter count and results could not be independently verified until the weights shipped ^[5][6]. Unite.AI reports the weights and technical report went up on Hugging Face and GitHub on July 27 under a custom "Kimi K3 License" that reads like MIT for most of its length but requires a separate agreement for model-as-a-service use above $20 million in revenue over 12 months ^[7]. Read the license file yourself before building on it.

What changed for buyers

Three practical points follow from the table and the releases.

1. The unit of comparison is cost per completed task. Anthropic's own framing, and OpenAI's tiers, both assume you will match a model to a job. A 40x spread in output price between the top and bottom rows is larger than most accuracy gaps on routine work.

2. Effort is now a budget line. Settings like Opus 5's effort levels change token consumption per request, so the same model can have very different bills depending on configuration.

3. Prices move in both directions and mid-quarter. A cut of 80% on one tier arrived three weeks after launch. Budgets and forecasts should assume re-pricing.

Vendor benchmark charts remain a weak basis for purchasing. The reliable approach is a small evaluation set drawn from your own documents and workflows, run against two or three tiers, with cost recorded per task.

Implication for businesses: treat model choice as a routing decision, not a platform bet. Test the cheapest tier that passes your own acceptance checks, keep a stronger model for the hard cases, and revisit the mix every quarter as prices change.