Three frontier AI models landed within 15 days: OpenAI’s GPT-5.6 family on July 9, Moonshot AI’s Kimi K3 on July 16, and Anthropic’s Claude Opus 5 on July 24, 2026. Net effect: near-flagship performance now costs half what it did in June, and the strongest open-weight model ever goes downloadable July 27.
This guide compares pricing, benchmark claims, and deployment reality across all three, plus the open-weight field, so you can decide what to run before your next invoice. OpenAI now names its models like NASA probes and Anthropic like tiers of Greek epic; the invoices, at least, remain in plain English.
Quick Answer
- Claude Opus 5 ($5/$25 per million tokens) nearly matches Fable 5 at half its cost, per CNBC.
- GPT-5.6 Sol claims the top coding-efficiency spot; budget Luna beats Opus 4.8 at $1/$6, per TechCrunch.
- Kimi K3 is a 2.8-trillion-parameter open model; weights drop July 27, per VentureBeat.
- Self-hosters should shortlist GLM-5.2 (MIT license, GPQA Diamond 91.2%).
What just changed in the AI model market?
As of July 2026, the frontier tier has three fresh entrants and a price floor that dropped by roughly half in one month. Anthropic released Claude Fable 5 and its restricted twin Mythos 5 on June 9, then followed with the cheaper Opus 5 on July 24.
OpenAI answered on July 9 with the GPT-5.6 family in three sizes. Moonshot AI, the Beijing lab behind Kimi, shipped Kimi K3 on July 16 and promised full open weights by July 27.
The competitive pattern is new. Instead of one flagship per lab, each vendor now sells a ladder of models priced against rivals’ mid-tiers, and open-weight releases trail the frontier by weeks rather than years.
What does Claude Opus 5 offer, and what does it cost?
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens, half the price of Fable 5, while coming close to it on Anthropic’s published evaluations, per CNBC’s July 24 report.
It also beats Opus 4.8, its predecessor, on agentic coding, problem-solving, knowledge work, and computer use, according to MacRumors’ coverage of the announcement.
Anthropic made Opus 5 the default model on Claude Max, and it is now the strongest model available on the Pro plan. A fast mode runs at 2.5x speed for double the base price.
One caveat from the same coverage: Opus 5 approaches Mythos 5 at finding cybersecurity vulnerabilities but was not trained for exploitation tasks, where it remains substantially behind. For most teams that distinction will never matter; for security teams it is the whole decision.
How do GPT-5.6 Luna, Terra, and Sol compare?
GPT-5.6 splits into three models: Sol ($5/$30 per million tokens) for complex reasoning and coding, Terra ($2.50/$15) for everyday work, and Luna ($1/$6) for speed, per TechCrunch’s launch report. All three are live across ChatGPT, Codex, and the OpenAI API.
OpenAI’s claims center on efficiency. Sol scores 80 on the Artificial Analysis Coding Agent Index, edging Fable 5 by 2.8 points on less than half the output tokens, per the same TechCrunch coverage.
OpenAI calls that 54% more token efficient on coding tasks. Terra lands just above Fable 5 on the same index; Luna outperforms Opus 4.8.
OpenAI also calls GPT-5.6 its strongest cybersecurity model yet, positioned for threat modeling, code review, and blue-team work. The launch followed a delayed rollout: a trusted-partner preview opened June 26 after government restrictions pushed back the public date.
Why does Kimi K3 matter for open source AI?
Kimi K3 is the largest open-weight model ever announced: 2.8 trillion parameters with a 1,048,576-token context window, per VentureBeat. Moonshot released API access on July 16 and scheduled the full weights for July 27, 2026.
“The largest open-weight AI model ever,” as Tom’s Hardware put it, and one that beats Claude Fable 5 on the Frontend Code Arena, where K3 holds the #1 spot at 1,679 points.
Independent numbers back the frontier claim. On Artificial Analysis’ long-horizon knowledge-work evaluation, K3 scored a 1,547 Elo, a 732-point jump over Kimi K2.6, trailing only Fable 5, per Simon Willison’s hands-on write-up.
On the GDPval-AA v2 occupational benchmark it placed third at 1,687, behind Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8), per VentureBeat.
Pricing tells you Moonshot knows what it built. At $3/$15 per million tokens, K3 costs about triple K2.6 ($0.95/$4) and is the most expensive API a Chinese lab has shipped, per Willison, whose single test prompt burned 13,241 reasoning tokens and about 25 cents.
Which open-weight models can you actually self-host?
For teams with real hardware budgets, four open-weight families matter as of July 2026, per Thunder Compute’s July roundup. The weights are free; the eight H100s are not.
| Model | Size (active) | License | Standout number |
|---|---|---|---|
| GLM-5.2 | 744B (40B) | MIT | GPQA Diamond 91.2%, Terminal-Bench 81.0 |
| Kimi K2.5 | 1T (32B) | Modified MIT | SWE-Bench 76.8%, native multimodal |
| DeepSeek-R1 | 671B (37B) | MIT | Distilled variants down to 1.5B |
| Llama 4 Maverick | 400B (17B) | Llama 4 Community | 1M-token context window |
GLM-5.2 is the reasoning leader of the group but needs roughly 245GB of memory even at 2-bit quantization. DeepSeek-R1’s distills are the realistic consumer-GPU path. Kimi K3 joins this table July 27, though at 2.8T parameters it is a datacenter artifact, not a workstation download.
Which model should you choose for your workload?
Match the model to the job and the invoice, not the leaderboard. The price-performance table as of July 25, 2026:
| Model | Maker | Price in/out (per M) | Best documented claim |
|---|---|---|---|
| Claude Opus 5 | Anthropic | $5 / $25 | Near-Fable 5 at half cost (CNBC) |
| GPT-5.6 Sol | OpenAI | $5 / $30 | 80 on AA Coding Agent Index (TechCrunch) |
| GPT-5.6 Luna | OpenAI | $1 / $6 | Outperforms Opus 4.8 (TechCrunch) |
| Kimi K3 | Moonshot AI | $3 / $15 | #1 Frontend Code Arena (Tom’s Hardware) |
Coding agents: run Sol and Opus 5 head-to-head on your own repo; the vendors’ claims point in opposite directions, which is exactly when your own eval matters. Our June 2026 coding agent rankings cover the harness side of that stack.
High-volume drafting and extraction: Luna at $1/$6 is the price anchor of the month. K2-class open models undercut it only if your GPUs are already paid for.
Frontend-heavy work: Kimi K3’s arena lead makes it worth a trial through the API now, weights later. Routing a two-week trial through a gateway keeps the switch reversible; we tested that pattern in our Portkey AI gateway review.
Security teams: both labs now market cyber capability explicitly. Read the fine print on what each model will refuse; capability and permission are different columns in the spec sheet.
Key Takeaways
- Frontier pricing halved in July 2026: Opus 5 at $5/$25, Sol at $5/$30, Luna at $1/$6.
- Kimi K3’s July 27 weights release puts a frontier-adjacent model in open hands for the first time.
- Benchmark claims now conflict head-on between vendors; your own 20-prompt eval decides.
- Self-hosting only pays at sustained volume; GLM-5.2 and DeepSeek-R1 distills lead that path.
FAQ
Is Claude Opus 5 better than Fable 5?
No. Anthropic positions Opus 5 as nearly matching Fable 5 at half the price, per CNBC. Fable 5 remains the top-scoring model on most published evaluations; Opus 5 is the value pick.
When do Kimi K3’s open weights come out?
Moonshot AI scheduled the full weight release for July 27, 2026, per VentureBeat. API access has been live since July 16 at $3/$15 per million tokens.
What is the cheapest of the new frontier models?
GPT-5.6 Luna, at $1 per million input tokens and $6 per million output tokens, per TechCrunch. OpenAI states it outperforms Claude Opus 4.8, last generation’s flagship tier.
Can I run Kimi K3 on my own hardware?
Realistically no. At 2.8 trillion parameters it needs datacenter-class multi-GPU hardware. GLM-5.2 (roughly 245GB memory at 2-bit quantization) or DeepSeek-R1 distills are the practical self-host choices, per Thunder Compute.
Do these releases change anything for ChatGPT or Claude subscribers?
Yes. Opus 5 is now the default on Claude Max and the strongest model on Claude Pro, and GPT-5.6 is live across ChatGPT plans, so both subscriptions got materially stronger without a price change.
Sources: Anthropic, TechCrunch, Simon Willison; CNBC, MacRumors, VentureBeat, Tom’s Hardware, and Thunder Compute cited inline.
Last updated: 2026-07-25
