The AI model menu in 2026: when to use Sol, Terra, Luna, Grok 4.5, Sonnet 5, and Muse Spark 1.1

July 28, 202612 minAI, GPT-5.6, Grok 4.5, Claude Sonnet 5, Muse Spark 1.1, Routing

Short answer (60 seconds): Between June 30 and July 15, 2026 there was a flood of frontier models. Seven launches in nine days (July 8-9): Grok 4.5 (xAI), GPT-Live-1 + mini (OpenAI full-duplex voice), GPT-5.6 Sol/Terra/Luna (OpenAI) and Muse Spark 1.1 (Meta). Before that, on June 30, Anthropic had launched Claude Sonnet 5. The question stopped being "OpenAI or Anthropic" — it's now "which of the six+ models for which use case?". This is the cheatsheet with pricing, cost-per-task, and the caveats that change your routing matrix (EU availability, Anthropic's 30-day retention, Sonnet 5's expiring intro pricing, Claude's new tokenizer).

The frontier menu of July 2026 fits on a single page — and that's what makes this moment strange. Five labs (Anthropic, OpenAI, xAI, Meta, and off-stage Google with no release) moved the catalog in less than a month. Every time you looked at a table, a cell changed.

For SaaS founders and CTOs the decision is no longer "which vendor" — it's "which model from which vendor for which workload". The sticker price stopped being enough: on coding agent, GPT-5.6 Luna runs ~USD 0.21 per task on the Artificial Analysis Coding Agent Index; Claude Fable 5 runs ~USD 11.80 — 56x for the same task. If you're not routing by cost-per-task, you're paying the bill of a lab that isn't right for you.

This is the cheatsheet I had in my Notion, sorted so you can open it when you need to decide. One table with everything, then one paragraph per model with what matters, then at the end the routing rules I apply with clients.

The table: the seven new models of the cycle

*Sonnet 5 intro pricing runs through August 31, 2026.

Verified sources: xAI, OpenAI API pricing, Anthropic, Meta AI, OpenAI GPT-Live. Compiled listing from Capital & Compute.

The chart that matters: cost-per-task, not per token

The sticker price misleads. What changes your invoice is what a finished task costs you, multiplying the actual token consumption by the model's rate. On the Artificial Analysis Coding Agent Index, a typical agentic coding task runs:

The gap between the cheapest and the most expensive is ~56x for the same coding agent task. This is the central argument for routing instead of picking one model.

Per model: what matters about each one

Claude Sonnet 5 (Anthropic, Jun 30)

Anthropic's mid-tier, shipped as the "default" for agents running on Claude Code. Intro pricing $2 input / $10 output per 1M tokens through August 31, 2026; afterward it jumps to $3 / $15. 1M token context, max output 128k.

What stands out: the intro pricing is an evaluation window of a month and a half. If you're already on Anthropic SDK, there's no reason to switch now — capture the discount. Important caveat: Sonnet 5 uses a new tokenizer that produces ~30% more tokens for the same text, so the list price underestimates the real cost when comparing byte-for-byte with GPT-5.6.

GPT-5.6 Sol (OpenAI, Jul 9 — flagship)

$5 input / $30 output per 1M tokens. The flagship of the 5.6 family relaunch. Has ultra mode (parallel subagents), max mode (reasoning effort at the ceiling), and Programmatic Tool Calling in GA.

What stands out: leads the AA Coding Agent Index (80 vs Fable 5 at 77.2) and "Agents' Last Exam" (53.6 vs Fable 5 at 40.5). On SWE-Bench Pro it trails Fable 5: 64.6% vs 80.4%. It's the pick for long agentic tasks where you need sustained coordination. Caveat: ultra mode burns many more tokens per task — it belongs in the "quality-escalation" column, never "savings".

GPT-5.6 Terra (OpenAI, Jul 9 — mid)

$2.50 / $15 per 1M tokens. Positioned by OpenAI as "GPT-5.5 quality at half the price". Available in Codex and the API.

What stands out: the menu's sweet spot. If your workload is drafting long documents, generating non-extreme application code, or knowledge work where Fable 5 is overkill, Terra is the balanced-cost option. On the AA Coding Agent Index it runs $0.55 per task — half of Sol with comparable quality on most marketing and back-office workloads. If you had GPT-5.5 integrated, this is the obvious migration.

GPT-5.6 Luna (OpenAI, Jul 9 — volume)

$1 input / $6 output per 1M tokens. The family's volume tier, available on the free tier in addition to paid plans.

What stands out: the cheapest by sticker. On the AA Coding Agent Index it runs ~$0.21 per task — the market floor. Caveat: the same task that costs $0.21 on Luna might cost less on Muse Spark if the load is output-heavy (Luna charges $6 output, Muse Spark $4.25). For input-dominated workloads (classification, extraction, RAG retrieval scoring), Luna is the right call. For agentic loops with heavy output, Muse Spark 1.1 wins below.

Grok 4.5 (xAI, Jul 8)

$2 input / $6 output per 1M tokens. 500K context, ~80 tokens/second output, co-trained with Cursor on agent loop traces.

What stands out: the cheapest frontier sticker combined with Cursor-style latency. If your IDE is Cursor, Grok 4.5 is tuned to the exact loop you run every day. Artificial Analysis measures it fourth on the Intelligence Index (54, behind Fable 5 at 60, Opus 4.8 at 56 and GPT-5.5 at 55). Caveat: not yet available in the EU. If your team is in Europe, wait until mid-July (xAI's dates) or use Cursor with another family in the meantime.

Muse Spark 1.1 (Meta, Jul 9)

$1.25 input / $4.25 output per 1M tokens. 1M context. Meta's first paid model (until now all of its work was open-weights). US-only preview for now, with free credits at launch.

What stands out: the cycle's surprise. Its output rate ($4.25/M) undercuts GPT-5.6 Luna ($6) and Claude Haiku 4.5 ($5), which is where the real savings live for output-heavy agentic loops. Leads the tool-use evals Meta published, but trails the frontier on pure coding — the numbers are vendor-reported, not independently audited yet. Caveat: US-only preview limits global adopters; capacity looks tight at launch (the free credits suggest so).

GPT-Live-1 and GPT-Live-1 mini (OpenAI, Jul 8)

Full-duplex voice models launched July 8. GPT-Live-1 becomes the new default for ChatGPT Voice on paid plans, GPT-Live-1 mini for the free tier. No API pricing published yet — OpenAI has a signup form for notification.

What stands out: full-duplex means the model listens and speaks at the same time. You can interrupt it like a human conversation, it does live translation, there's no explicit turn. If you're building voice UX in 2026 (voice customer support, voice-enabled agents, accessibility), this is the most relevant change of the month even though it's not accessible via API yet. Caveat: latency and quality specifically in Latin American Spanish haven't been independently measured yet — that's where LATAM SaaS usually finds friction.

When to pick each one (routing rules)

These are the rules I use with clients when I build a routing matrix. They aren't the only correct ones — they're the ones that have produced the best results on real SaaS workloads.

Default coding workhorse —> GPT-5.6 Terra. GPT-5.5 capabilities at half the price ($2.50/$15). Covers 80% of production coding workloads without needing frontier.

Hardest tasks, price no object —> GPT-5.6 Sol (max or ultra). If your load is multi-step agentic, multi-hour sessions with sustained tool use, or reasoning over large codebases. You accept the $5/$30 sticker because the cost-per-task stays competitive.

High-volume output-heavy agents —> Muse Spark 1.1 (if you're in the US) or GPT-5.6 Luna (global). If your agentic loop generates a lot of output, Muse Spark at $4.25/M undercuts Luna at $6/M. If your loop is input-heavy (RAG retrieval, classification), Luna at $1/M input is the floor.

Cursor users —> Grok 4.5. The co-training with Cursor means the model is optimized for the exact Cursor loop. Try a side-by-side sprint — often the difference shows up in latency, not in output.

Anthropic stack —> Claude Sonnet 5 before August 31. If your codebase already runs on Anthropic SDK, don't migrate for the sake of migrating. Capture the $2/$10 intro discount, run your own evals, then decide.

Voice UX —> GPT-Live-1. Even though it's only via ChatGPT for now, the qualitative shift from half-duplex to full-duplex redefines voice UX. Wait for the public API but start designing for it.

The four asterisks that change your matrix

The menu looks clean in the table. In production there are four facts that change the decision for specific teams:

1. EU availability of Grok 4.5. Not yet available in the EU. If your team runs in Europe, every "Grok 4.5" in the matrix above carries an asterisk. xAI targets mid-July 2026 — review before committing.

2. Data retention on Mythos / Fable 5. Anthropic applies mandatory 30-day retention on Mythos-class models (which includes Fable 5), and Zero Data Retention agreements don't cover them. If you pass client strategy drafts or unpublished material through them, that's a routing input on its own.

3. Sonnet 5 intro pricing expires August 31. The intro discount ($2/$10, 33% below standard rate) is an evaluation window. From September 1 it jumps to $3/$15. If your plan is to move anything significant to Sonnet 5, the moment is now.

4. New Claude tokenizer. Anthropic confirmed Sonnet 5 uses a tokenizer that produces ~30% more tokens for the same text. The "$/MTok" comparison against GPT-5.6 underestimates Claude's real cost at equal text volume. Multiply by 1.3 before comparing.

The LATAM angle: what changes for you

Three observations from Buenos Aires, not San Francisco:

1. Pricing in USD with no regional variation. OpenAI, Anthropic and xAI charge the same in Buenos Aires as in California. Meta doesn't differentiate either. This levels the playing field against US-based competitors on API cost — you make the difference in how you use it, not in how much you pay per token.

2. Data residency does matter. Fable 5 / Mythos 5 with 30-day retention is not an option if you handle client data that contractually requires local processing or no retention. For those cases, Grok 4.5 (with the EU caveat), Muse Spark (US-only), or a low GPT-5.6 tier (Luna) with appropriate data agreements are alternatives. OpenAI offers Zero Data Retention as an option on the enterprise API — review case by case.

3. Latin American Spanish and LATAM-specific evals. None of these models has been independently measured on Río de la Plata Spanish workloads specifically (Argentine legal, Mexican commercial contracts, Chilean regulation). Before committing to a large migration, run evals with your real prompts — the model "best on benchmarks" isn't necessarily the best for your regional case.

Want to discuss which of the six+ makes sense for your workload before switching models or migrating code? There's a CTA at the end with a free 30-minute call.

Frequently asked questions (summarized)

What's the most important thing that changed in July 2026? Seven models in nine days, from four different labs, with prices overlapping in tiers where no overlap existed before.

Which is the cheapest? GPT-5.6 Luna ($1/$6 MTok). On cost-per-task, also Luna (~$0.21).

Which is best for sustained agentic coding? GPT-5.6 Sol leads the AA Coding Agent Index; Claude Fable 5 leads SWE-Bench Pro. For sustained coding, Sol — the sticker is cheaper and the difference is small.

Is Sonnet 5 still relevant after GPT-5.6? Yes, especially if your stack is already on Anthropic. Capture the intro pricing before August 31.

Why is Muse Spark interesting if it trails on coding? The output rate undercuts Luna and Haiku — key for output-heavy agent loops. And Meta being Meta, the next releases could close the gap fast.

What about Gemini? Google didn't release anything new in July. Gemini 3.5 Pro is "coming soon" with no date. Worth watching when it lands.

Frequently asked questions

What changed in the AI model menu in July 2026?

Seven models shipped from four labs in nine days (July 8 and 9): Grok 4.5 (xAI), GPT-Live full-duplex voice (OpenAI), GPT-5.6 Sol/Terra/Luna (OpenAI), Muse Spark 1.1 (Meta). Claude Sonnet 5 (Anthropic) had shipped on June 30 just before. The frontier menu stopped being "choose between two labs" and became "choose between six+ models based on use case and budget per task".

What's the cheapest model per task in July 2026?

By sticker price, GPT-5.6 Luna ($1 input / $6 output per 1M tokens). By real cost per finished task on the Artificial Analysis Coding Agent Index, GPT-5.6 Luna completes a typical coding agent task for ~USD 0.21. Grok 4.5 runs ~$2.49 per task — its cheap sticker ($2/$6) gets eaten up by more output tokens than Luna on agentic coding workloads.

When does Grok 4.5 beat the alternatives?

Grok 4.5 was co-trained with Cursor on agent loop traces, served at ~80 tokens/second, $2/$6 MTok, 500K context. It's the right pick when your IDE is Cursor, when you need low response latency for real-time UI, or when you want the cheapest frontier sticker. Caveat: it isn't yet available in the EU.

GPT-5.6 Sol or Claude Fable 5 for agentic coding?

There's no single champion. On OpenAI's own GA tables, Claude Fable 5 leads SWE-Bench Pro (80% vs Sol 64.6%) and GDPval-AA Elo. Sol leads the AA Coding Agent Index (80 vs 77.2) and "Agents' Last Exam" (53.6 vs 40.5). On real workloads the gap is small and cost-per-task differs ~10x: Sol ~$1.04, Fable 5 ~$11.80. For sustained agentic coding, Sol is the current sweet spot; for sustained reasoning and edge cases, Fable 5.

What's special about Muse Spark 1.1?

It's Meta's first paid model. $1.25 input / $4.25 output per 1M tokens, 1M context, agentic + multimodal focus, US-only preview for now. Its output rate ($4.25) undercuts GPT-5.6 Luna ($6) and Claude Haiku 4.5 ($5), which is where the real savings live for output-heavy agentic workloads.

Claude Sonnet 5 or the new GPT-5.6 tiers?

Sonnet 5 stays the obvious pick if your stack is already on Anthropic: $2/$10 intro pricing through August 31, 1M context, the same SDK you're already using. The alternative is GPT-5.6 Terra ($2.50/$15) at GPT-5.5 quality, half the price of GPT-5.5 — comparable on marketing drafting or long documents. If your workload is coding agents on Claude Code, stay on Sonnet 5 through August to capture the discount.

What's GPT-Live and why does it matter?

GPT-Live-1 and GPT-Live-1 mini are OpenAI's full-duplex voice models launched July 8: the model listens and speaks at the same time, you can interrupt it naturally, and it does live translation. GPT-Live-1 becomes the default ChatGPT Voice for paid plans; the mini version for free users. No API pricing published yet. If you're building voice UX in 2026, it's the most relevant change of the month.