Capital & Compute
· Updated August 14, 2026· ai· pricing· economics

New AI Models Released in July 2026: The Full List

Every new AI model released in July 2026 with dates, per-token prices and primary sources, plus the announced models still ahead on the release timeline.

By Capital & Compute

Eight new AI models shipped in the first two weeks of July 2026: Grok 4.5 from xAI (July 8), the GPT-5.6 family of Sol, Terra and Luna from OpenAI (July 9), Meta’s Muse Spark 1.1 (July 9), OpenAI’s two GPT-Live voice models (July 8), and Inkling, the first open-weight model from Mira Murati’s Thinking Machines (July 15). Every one of them is verified against a primary source below, with the per-token rate where the provider has published one. This page is the July installment of a monthly series; the next one covers the new AI models released in August 2026. The complete dated list, including everything that shipped in the second half of the month, lives in the record of AI model releases by month; current per-token prices and upcoming models are in the AI model tracker.

Every AI model released in July 2026

Model Maker Released Price (in/out per Mtok) What it is
Grok 4.5 xAI Jul 8 $2 / $6 Frontier model, coding-agent focus, 500K context
GPT-Live-1 OpenAI Jul 8 Not published Full-duplex voice model for ChatGPT paid tiers
GPT-Live-1 mini OpenAI Jul 8 Not published Smaller voice model, default for free ChatGPT Voice
GPT-5.6 Sol OpenAI Jul 9 $5 / $30 Flagship of the 5.6 suite, max and ultra reasoning modes
GPT-5.6 Terra OpenAI Jul 9 $2 / $12 Mid tier, pitched at GPT-5.5 performance for half the cost; cut 20% Jul 30
GPT-5.6 Luna OpenAI Jul 9 $0.20 / $1.20 Budget tier for routine and high-volume work; cut 80% Jul 30
Muse Spark 1.1 Meta Jul 9 $1.25 / $4.25 Agentic multimodal model, first paid Meta model, 1M context
Inkling Thinking Machines Jul 15 $1.87 / $4.68 First open-weight model from Mira Murati’s lab, 975B/41B MoE, Apache 2.0
Claude Opus 5 Anthropic Jul 24 $5 / $25 New Anthropic flagship, effort dial, 1M context, replaces Opus 4.8

Two days, three labs, and a clear pattern: everyone shipped into the same week. OpenAI had previewed GPT-5.6 on June 26 behind a US-government evaluation gate, xAI and Meta both landed the day before and the day of its general availability. Nobody wants to launch into someone else’s news cycle, and in July nobody managed to avoid it. Inkling arrived a week later and outside that cluster, but it is the month’s other genuine event: not a price move by an incumbent, but a new lab entering the open-weight field for the first time.

Grok 4.5: the cheap frontier bid (July 8)

xAI released Grok 4.5 on July 8, 2026 at $2 input and $6 output per million tokens, with a 500K context window. The model was trained jointly with Cursor on coding-agent traces, and the positioning is unambiguous: a frontier-class coder at a quarter of flagship pricing.

The benchmark picture needs the vendor-versus-neutral caveat. On the xAI-run DeepSWE 1.0, Grok 4.5 leads Claude Opus 4.8 (62.0 vs 55.75); on the neutral DeepSWE 1.1 it trails (53 vs 59), and it trails on SWE-Bench Pro too (64.7 vs 69.2). Artificial Analysis, an independent evaluator, places it fourth on its Intelligence Index at 54, behind Claude Fable 5 (60), Opus 4.8 (56) and GPT-5.5 (55). Fourth place at $2/$6 is the actual pitch. We ran the full cost-per-task breakdown on Grok 4.5 the day it landed.

One more detail worth watching if you use Cursor: the joint training arrangement means Grok 4.5 is tuned for the exact agent loop Cursor runs. It is also one of the faster frontier models: xAI quotes around 80 output tokens per second and Artificial Analysis measured 91.3, and speed compounds in agentic work because every loop iteration waits on the model.

GPT-5.6 Sol, Terra and Luna: the family goes GA (July 9)

OpenAI made the GPT-5.6 suite generally available on July 9 across ChatGPT, Codex and the API, ending the two-week government-gated preview that began June 26. They launched at $5/$30 (Sol), $2.50/$15 (Terra) and $1/$6 (Luna), each with cached input at 90% off. On July 30 OpenAI cut Terra 20 percent and Luna 80 percent, so the official API pricing page now lists Sol at $5/$30, Terra at $2/$12 and Luna at $0.20/$1.20.

Sol is the headline. With its ultra mode (parallel subagents) it scores 91.91% on Terminal-Bench 2.1, ahead of Claude Mythos 5 at 88% and GPT-5.5 at 83.4%. Terra looked like the model most teams should price first, at about $0.55 per task on the Artificial Analysis Coding Agent Index against Sol’s $1.04. The July 30 cut changed that answer: Terra now runs about $0.44 per task and Luna about $0.04, a 10x gap for two points of coding composite, so Luna is the tier to price first. The launch-day analysis of Sol, Terra and Luna covers what changed since the preview, and the GPT-5.6 pricing deep dive has the per-task math.

Cost per coding-agent task: July 2026 releases vs incumbentsHorizontal bar chart of cost per task on the Artificial Analysis Coding Agent Index at July 31 2026 rates. GPT-5.6 Luna costs about $0.04 per task, GPT-5.6 Terra $0.44, GPT-5.6 Sol $1.04, Grok 4.5 $2.49, GPT-5.5 $5.07, and Claude Fable 5 $11.80. The three GPT-5.6 tiers and Grok 4.5, all July releases, occupy the cheap end; the incumbents GPT-5.5 and Claude Fable 5 anchor the expensive end.$0.00$2.00$4.00$6.00$8.00$10.00$12.00GPT-5.6 Luna (Jul 9)$0.04GPT-5.6 Terra (Jul 9)$0.44GPT-5.6 Sol (Jul 9)$1.04Grok 4.5 (Jul 8)$2.49GPT-5.5 (Apr 2026)$5.07Claude Fable 5 (Jun 2026)$11.80
Cost per coding-agent task: July 2026 releases vs incumbents
ItemValue
GPT-5.6 Luna (Jul 9)$0.04
GPT-5.6 Terra (Jul 9)$0.44
GPT-5.6 Sol (Jul 9)$1.04
Grok 4.5 (Jul 8)$2.49
GPT-5.5 (Apr 2026)$5.07
Claude Fable 5 (Jun 2026)$11.80
What one coding-agent task costs on the Artificial Analysis Coding Agent Index, as of July 31, 2026. Each bar is a model running in its native harness (GPT-5.6 tiers in Codex, Grok 4.5 in Grok Build, GPT-5.5 in Codex, Claude Fable 5 in Claude Code), so the figures compare agent stacks, not raw models. July arrivals in color. The Terra and Luna bars apply the July 30 price cut to the token counts Artificial Analysis measured, since Artificial Analysis had not yet republished its own per-task figures.Source: Artificial Analysis Coding Agent Index token counts, repriced at the July 30 2026 rates in the Capital & Compute model registry

That chart is the month in one picture. Four of the six cheapest slots belong to models that did not exist on July 1. The gap between the cheapest July arrival and the most expensive incumbent is roughly 280x per task after the July 30 cut, which is why sticker-rate comparisons keep misleading people; if per-task economics is the decision you are making, the coding-agent cost-per-task tracker keeps these figures current.

Muse Spark 1.1: Meta starts charging (July 9)

Meta released Muse Spark 1.1 on July 9, its first model ever to carry a price: $1.25 input and $4.25 output per million tokens, with $20 in free credits and a 1 million-token context window, in a US-only API preview. On the launch benchmarks Meta published, it leads every tool-use test and trails the frontier on every pure-coding one; the numbers are vendor-reported and not independently reproduced. The output rate mattered more than it looked, because agentic work is output-heavy, and $4.25 undercut both GPT-5.6 Luna ($6 at the time) and Claude Haiku 4.5 ($5). OpenAI’s July 30 cut took Luna to $1.20 and erased that edge. Our full verdict on Muse Spark 1.1 covers where the pitch holds and where it does not.

Inkling: Thinking Machines’ first open-weight model (July 15)

Mira Murati’s Thinking Machines released Inkling on July 15, a 975 billion-parameter (41 billion active) mixture-of-experts model released under a full Apache 2.0 license, with the weights and an NVFP4 quantized variant published on Hugging Face. It is the company’s first model release since it went public with the Tinker fine-tuning platform, and its first to ship open weights rather than stay behind an API. The company’s own framing is unusually blunt for a launch post: Inkling is “not the strongest overall model available today, open or closed.” The pitch instead is a broad, multimodal (text, image, audio) base built to be fine-tuned on Tinker, not a leaderboard-topping flagship.

The benchmark picture matches that framing. On Thinking Machines’ own comprehensive benchmark table, Inkling edges DeepSeek V4 Pro on AIME 2026 (97.1 vs 96.7) and is close on SWE-Bench Verified (77.6 vs 80.6), but it trails GLM-5.2 and Kimi K2.6 substantially on Terminal Bench 2.1 (63.8 vs 82.7 and 71.3) and trails DeepSeek V4 Pro on SimpleQA Verified (43.9 vs 57.0). Where it does lead is safety: a FORTRESS Adversarial score of 78.0, ahead of Nemotron 3 Ultra, GLM-5.2 and DeepSeek V4 Pro, plus a rare positive Artificial Analysis Omniscience factuality score.

Pricing runs through Thinking Machines’ own Tinker platform, not a third-party API: $1.87 per million input tokens and $4.68 per million output at the 64K-context tier (the 256K tier costs exactly double), with cached input at 80% off. That is a limited-time launch rate. Tinker’s own pricing docs show prefill and sample prices rising roughly 50% on July 17, two days after launch. The full architecture, benchmark, and licensing breakdown is in the dedicated Inkling post.

GPT-Live-1 and GPT-Live-1 mini: the quiet release (July 8)

The July release most roundups skip: OpenAI shipped two full-duplex voice models on July 8, announced in Introducing GPT-Live on the OpenAI site. Full-duplex means the model listens and speaks at the same time, so you can interrupt it naturally and it can do live translation. GPT-Live-1 becomes the default voice model for paid ChatGPT tiers, the mini version for free users. No API access yet and no published per-token rate; OpenAI has a sign-up form for API notification. We list them because a release tracker that only counts text models undercounts the month.

Did Anthropic release a new model in July 2026?

Yes. On July 24, 2026 Anthropic shipped Claude Opus 5, its new flagship, at $5/$25 per million tokens, the same rate as the Opus 4.8 it replaces. It adds an effort dial (low to max) that trades intelligence against token spend, and on the independent Artificial Analysis Intelligence Index it edges the far pricier Fable 5, 61 to 60 at max effort. The Opus 5 pricing and benchmarks breakdown has the full picture. Earlier in the year Anthropic shipped Claude Sonnet 5 on June 30 and Claude Fable 5 and Mythos 5 on June 9.

The June 30 release looked like the one with a clock on it. Claude Sonnet 5 launched as the mid-tier Claude with introductory pricing of $2/$10 per million tokens through August 31, 2026, after which it was to revert to $3/$15. That deadline is gone: Anthropic’s pricing page now states the $2/$10 rate is the standard price and the September 1 increase will not occur, so the discount became permanent rather than expiring (updated August 14, 2026). It carries a 1 million-token context window and 128k max output. The Sonnet 5 pricing and benchmarks breakdown has the full picture.

Above it sits Claude Fable 5 at $10/$50 per Mtok, the top generally available capability tier, though Opus 5 now matches it on the independent Intelligence Index at half the token price. We track every confirmed date in the release tracker.

Did Google release a new model in July 2026?

Also no. Google’s most recent model release is Gemini 3.5 Flash on May 19, 2026. The current flagship remains Gemini 3.1 Pro.

The interesting part is what Google has queued. The official Gemini models page lists Gemini 3.5 Pro as coming soon, with no date. That model completes the 3.5 series that Flash started, and it is the most likely next frontier event on the calendar: Google is the only top-three lab that has not shipped a frontier-tier model since May, while OpenAI, Anthropic, xAI and Meta all have. If the July pattern holds (labs shipping into each other’s launch windows), expect 3.5 Pro to land near whatever OpenAI or Anthropic does next, not in a quiet week.

The 2026 release timeline so far

For the “when did each model come out” question, here is the release sequence since April, every date grounded to the model registry and its primary sources:

When Model Maker Why it mattered
April 2026 GPT-5.5 OpenAI The flagship many expected to be GPT-6; held the top coding spots until July
April 2026 DeepSeek V4 (preview) DeepSeek 1M context open-weights flagship, Pro and Flash tiers
May 19 Gemini 3.5 Flash Google Started the 3.5 series; still Google’s newest release
May 2026 Qwen3.7 Max Alibaba Alibaba Cloud flagship, aggressive API pricing
June 9 Claude Fable 5 + Mythos 5 Anthropic New top tier at $10/$50; Mythos 5 gated to approved orgs
June 9 North Mini Code Cohere Small coding model aimed at enterprise deployment
June 2026 GLM-5.2 Zhipu Open-weights flagship, strong long-horizon coding claims
June 2026 Kimi K2.7 Code Moonshot Coding-tuned K2 line update with a high-speed variant
June 30 Claude Sonnet 5 Anthropic Mid-tier refresh at $2/$10, now the permanent rate
July 8 Grok 4.5 xAI Frontier bid at $2/$6, Cursor co-training
July 8 GPT-Live-1 + mini OpenAI Full-duplex voice, ChatGPT Voice default
July 9 GPT-5.6 Sol, Terra, Luna OpenAI New flagship suite, three price tiers
July 9 Muse Spark 1.1 Meta First paid Meta model, agentic focus
July 15 Inkling Thinking Machines First open-weight model from a new frontier lab; Apache 2.0
July 24 Claude Opus 5 Anthropic New flagship at $5/$25 with an effort dial; matches Fable 5 on the AA Index

Read the cadence, not just the rows. Three months, five labs, and at least one significant release every two to three weeks. This is the environment our thesis on why the 2026 model race keeps inverting price expectations describes: capability leapfrogs every few weeks, so the premium you pay for “the best model” has a shelf life measured in weeks, and per-task cost matters more than the leaderboard position of the moment.

Which July model should you actually use?

Depends on what you are optimizing. The honest short answers, based on the verified rates and the independent numbers above:

  • Default coding workhorse: GPT-5.6 Luna, as of July 30. It runs about $0.04 per task against Terra’s $0.44 for two fewer points of coding composite, so Terra now has to earn a 10x premium on your own evals. Terra remains the safer default only if your tasks are hard enough that the gap shows up.
  • Hardest tasks, price no object: GPT-5.6 Sol in max or ultra mode, or Claude Fable 5 if your stack is already on Claude Code. Sol currently holds the Terminal-Bench 2.1 top score; Fable 5 still wins SWE-Bench Pro. Put your own task mix into the cost-per-task calculator before deciding, because the two differ by roughly 10x per task.
  • High-volume, cost-sensitive agents: GPT-5.6 Luna, and it is no longer close. After the July 30 cut Luna is cheaper than Muse Spark 1.1 on both input ($0.20 against $1.25) and output ($1.20 against $4.25), so the output-rate argument that favoured Muse Spark at launch no longer applies.
  • Cursor users: Grok 4.5 deserves a trial run. It was co-trained on Cursor agent traces and priced at $2/$6 to be an easy switch.
  • Anthropic-stack teams: Claude Sonnet 5 at $2/$10 is no longer a deadline. Anthropic cancelled the September 1 increase and made the rate standard, so there is no longer a reason to rush an eval.
  • A permissive-license fine-tuning base: Inkling. Apache 2.0, frontier-scale (975B/41B), and self-hostable, though it is not benchmark-leading out of the box; the pitch is customization on Tinker, not raw capability.

For a head-to-head on your own workload, the model comparison tool prices two or three models against the same task, and the value leaderboard ranks the whole field by benchmark points per dollar.

The release timeline from here

What is confirmed, announced, or credibly rumored as of July 11, 2026:

  1. Gemini 3.5 Pro (Google DeepMind): announced, no date. The official Gemini models page lists 3.5 Pro as coming soon. It completes the 3.5 series that started with Gemini 3.5 Flash in May, and it is the obvious next frontier event.
  2. GPT-6 (OpenAI): rumor only. No model page, no date, no announcement. OpenAI has said a new frontier model is in training, and the independent LifeArchitect GPT-6 tracker collects the public signals; prediction markets lean toward a second-half-2026 launch. Treat every “GPT-6 pricing” article as fiction until OpenAI publishes a page.
  3. The pattern to expect: June and July each delivered a multi-model cluster (Claude Fable 5 and Mythos 5 on June 9, Sonnet 5 on June 30, then the July week above). Labs are shipping into each other’s launch windows, so the next big release will probably not arrive alone.

We update the release tracker as each of these moves from rumor to announcement to shipped, with the per-token rate verified against the provider’s own page before it enters the value leaderboard.

Frequently asked questions

What new AI models came out in July 2026?
Seven models so far: Grok 4.5 from xAI on July 8, the GPT-5.6 family (Sol, Terra, Luna) generally available from OpenAI on July 9, Muse Spark 1.1 from Meta on July 9, and the GPT-Live-1 and GPT-Live-1 mini voice models from OpenAI on July 8.
Did OpenAI release a new model in July 2026?
Yes, twice. On July 8 it shipped GPT-Live-1 and GPT-Live-1 mini, full-duplex voice models for ChatGPT Voice. On July 9 it made the GPT-5.6 suite generally available across ChatGPT, Codex and the API after a government-gated preview that started June 26, then on July 30 it cut two of the three tiers. Current rates are Sol at $5/$30 per million tokens, Terra at $2/$12 and Luna at $0.20/$1.20.
Did Anthropic release a new model in July 2026?
No. As of July 11, 2026, Anthropic has not released a model in July. Its newest models are Claude Sonnet 5 (June 30, at $2/$10 per Mtok, a rate launched as introductory and since made permanent) and Claude Fable 5 and Mythos 5 (June 9). Anthropic went on to release Claude Opus 5 on July 24.
What is the latest AI model right now?
As of July 15, 2026, the newest release is Inkling, the first open-weight model from Mira Murati's Thinking Machines. Before that, the newest generally available models were GPT-5.6 Sol, Terra and Luna (OpenAI) and Muse Spark 1.1 (Meta), both released July 9, one day after Grok 4.5 (xAI) and the GPT-Live voice models (OpenAI) on July 8.
What is Inkling?
Inkling is Thinking Machines' first open-weight model, released July 15, 2026: a 975 billion-parameter mixture-of-experts model (41 billion active) under an Apache 2.0 license, priced on the company's own Tinker platform at $1.87/$4.68 per million tokens. The company frames it as a customizable fine-tuning base rather than a frontier leader. Full breakdown in the dedicated [Inkling launch post](/blog/inkling-thinking-machines-launch/).
What is the cheapest new AI model in July 2026?
GPT-5.6 Luna, at $0.20 input and $1.20 output per million tokens after OpenAI cut it 80 percent on July 30, 2026. It launched at $1/$6, which Muse Spark 1.1 ($1.25/$4.25) could undercut on output; the cut ended that, and Luna is now cheaper on both sides. Applying the cut to Artificial Analysis token counts, Luna completes a coding-agent task for about $0.04.
What AI models are coming next after July 2026?
Gemini 3.5 Pro is the only announced one: Google DeepMind lists it as coming soon with no date. GPT-6 is rumored but unannounced; OpenAI has said a new frontier model is in training, and prediction markets lean toward a second-half-2026 launch.
When will GPT-6 be released?
There is no announced date. OpenAI has not published a GPT-6 model page or confirmed the name; it has only said a new frontier model is in training. Prediction markets currently favor a second-half-2026 launch. Any article quoting GPT-6 specs or pricing today is speculation.

Sources

Get each breakdown before it makes the rounds

You get one email when a new source-backed analysis goes live: what AI agents actually cost, which models are worth running, and what the benchmarks really mean. No hype.

No spam. Unsubscribe anytime.

← Back to Models & benchmarks