Last updated: August 2026
Grok 4.6, GPT-5.6 Sol, and Claude Fable 5 are the three frontier models in play in August 2026, and they sit within about two points of each other on published intelligence indexes. The gap that is not close is price: Grok 4.6 is $2/$6 per million tokens, GPT-5.6 Sol is $5/$30, and Claude Fable 5 is $10/$50.
TL;DR
- Price (per 1M tokens): Grok 4.6 is $2 in / $6 out (xAI). GPT-5.6 Sol is $5 in / $30 out (OpenAI). Claude Fable 5 is $10 in / $50 out (Anthropic). Grok output is 80% cheaper than Sol and 88% cheaper than Fable 5.
- Intelligence: On OpenAI’s own Artificial Analysis Intelligence Index v4.1, Fable 5 scores 59.9 and Sol 58.9. On xAI’s version of the same index, Grok 4.6 and Sol both score 61 and Fable 5 scores 62. The two tables use different settings and are not directly comparable.
- Coding: By OpenAI’s numbers, Fable 5 leads SWE-Bench Pro (80% vs Sol’s 64.6%), while Sol leads the Coding Agent Index and Terminal-Bench. By xAI’s numbers, Sol Max leads DeepSWE and Terminal-Bench while Grok 4.6 leads CursorBench and FrontierCode.
- What’s new: Grok 4.6 adds long-horizon agent focus at Grok 4.5’s price; GPT-5.6 splits into Sol/Terra/Luna tiers with new max and ultra settings; Fable 5 is Anthropic’s first public Mythos-class model, above the Opus line.
- Caveat: each maker reports on different eval versions and reasoning settings, so cross-vendor benchmark numbers do not line up. Read each table on its own terms.
All figures below come from the makers’ own pages: xAI on Grok 4.6, OpenAI on GPT-5.6, and Anthropic on Claude Fable 5.
Which is cheapest, Grok 4.6, GPT-5.6 Sol, or Claude Fable 5?
Grok 4.6 is by far the cheapest, and Claude Fable 5 is the most expensive. Grok 4.6 lists at $2 per million input tokens and $6 output; GPT-5.6 Sol at $5 and $30; Claude Fable 5 at $10 and $50. Against Fable 5, Grok 4.6 is 80% cheaper on input and 88% cheaper on output.
| Model (maker) | Input (per 1M) | Output (per 1M) | Cached input | Source |
|---|---|---|---|---|
| Grok 4.6 (xAI) | $2.00 | $6.00 | Not published; confirm on xAI’s docs | xAI |
| GPT-5.6 Sol (OpenAI) | $5.00 | $30.00 | 90% discount on cache reads | OpenAI |
| Claude Fable 5 (Anthropic) | $10.00 | $50.00 | 90% discount for prompt caching | Anthropic |
Takeaway: for output-heavy or long-running agent work, where output tokens dominate the bill, Grok 4.6 is the clear cost pick and Fable 5 carries a large premium that only pays off if its results justify roughly 5x to 8x the output cost. (Prices as of August 2026; Sol and Fable 5 both apply a 90% cached-input discount, and OpenAI bills cache writes at 1.25x the uncached input rate.)
Grok 4.6 vs GPT-5.6 Sol vs Claude Fable 5: specs at a glance
The three are close on headline capability and far apart on price. Context-window figures are not stated in plain text on every maker page, so the two flagged below should be confirmed against developer docs before you rely on them.
| Spec | Grok 4.6 | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|---|
| Maker | xAI | OpenAI | Anthropic |
| Released | August 12, 2026 | July 9, 2026 (GA) | June 9, 2026 (restored July 1) |
| Input / output (per 1M) | $2 / $6 | $5 / $30 | $10 / $50 |
| Context window (tokens) | ~500,000 | ~1,000,000 | ~1,000,000 |
| Model ID | grok-4.6 | gpt-5.6-sol | claude-fable-5 |
| Notable setting | xhigh reasoning | max, ultra (4 parallel agents) | high-effort self-validation |
Takeaway: Sol and Fable 5 offer the larger context windows and Grok 4.6 offers the lowest price; all three target long-running agentic work rather than simple chat.
Is Grok 4.6 better than GPT-5.6 Sol and Claude Fable 5?
On intelligence, no single model is clearly ahead; the three land within about two points of each other, but the ranking flips depending on whose table you read. OpenAI’s own figures put Claude Fable 5 slightly ahead of GPT-5.6 Sol on the Artificial Analysis Intelligence Index v4.1 (59.9 vs 58.9). xAI’s figures put Grok 4.6 level with Sol at 61 and one point behind Fable 5 at 62. (OpenAI; xAI)
Why do the benchmark numbers differ between sources?
Because each lab runs different eval versions, harnesses, and reasoning settings, then reports its own results. GDPval is a clear example: OpenAI reports Fable 5 at 1,759.6 Elo and Sol at 1,747.8, while xAI reports Fable Max at 1,741, Sol Max at 1,728, and Grok 4.6 at 1,753. Same benchmark name, different runs, different numbers. Treat the two tables below as internally consistent but not cross-comparable, and wait for a neutral third party to run all three on one harness for a true head-to-head. No independent Artificial Analysis or LMSYS numbers covering all three models were published at the time of writing.
Grok 4.6 vs GPT-5.6 Sol vs Claude Fable 5: OpenAI’s benchmark table
By OpenAI’s own testing, GPT-5.6 Sol leads on coding-agent and terminal work while Claude Fable 5 leads on raw software resolution and intelligence. These are OpenAI-run figures, so read them as OpenAI’s methodology.
| Benchmark (unit) | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|
| AA Intelligence Index v4.1 (score) | 58.9 | 59.9 |
| AA Coding Agent Index v1.1 (score) | 80.0 | 77.2 |
| SWE-Bench Pro (%) | 64.6 | 80.0 |
| DeepSWE v1.1 (%) | 72.7 | 69.7 |
| Terminal-Bench 2.1 (%) | 88.8 (91.9 with ultra) | 83.1 |
| GDPval-AA v2 (Elo) | 1,747.8 | 1,759.6 |
| Agents’ Last Exam (%) | 52.7 | 40.5 |
| GPQA Diamond (%) | 94.6 | 92.6 |
| Toolathlon (%) | 58.0 | 61.7 |
Takeaway: OpenAI’s data shows Sol and Fable 5 trading the lead eval by eval, with Fable 5 notably stronger on SWE-Bench Pro and Sol notably stronger on long-horizon Agents’ Last Exam. Grok 4.6 is absent because OpenAI did not test it. (OpenAI)
Grok 4.6 vs GPT-5.6 Sol vs Claude Fable 5: xAI’s benchmark table
xAI’s launch table is the only one of the three that includes Grok 4.6. xAI states its competitor figures come from the other developers’ published system cards or leaderboards, and it ran these evals itself, so treat every row as self-reported by xAI.
| Benchmark (unit) | Grok 4.6 | GPT-5.6 Sol Max | Fable 5 Max |
|---|---|---|---|
| AA Intelligence Index (score) | 61 | 61 | 62 |
| GDPVal-AA v2 (score) | 1,753 | 1,728 | 1,741 |
| CursorBench v3.2 (%) | 69.9 | 67.2 | 70.5 |
| DeepSWE v1.1 (%) | 65.9 | 73.0 | 70.0 |
| FrontierCode v1.1 Extended (%) | 61.3 | 60.6 | 64.9 |
| APEX-Agents (%) | 57.5 | 56.7 | 59.2 |
| Terminal-Bench v3.0 (%) | 26.0 | 34.6 | 34.1 |
| AA-Briefcase (score) | 1,577 | 1,502 | 1,574 |
| Harvey LAB / Vals (%) | 15.8 | 2.5 | 11.3 |
Takeaway: by xAI’s numbers, Grok 4.6 leads on GDPVal, AA-Briefcase, and the Harvey legal eval, sits level with Sol on intelligence, and trails on DeepSWE and Terminal-Bench. Note the Terminal-Bench rows differ sharply from OpenAI’s table because xAI used a different, harder version (v3.0 vs 2.1), which is exactly why the two tables should not be merged. (xAI)
What is new in Grok 4.6?
Grok 4.6 keeps Grok 4.5’s price and adds long-horizon agent capability. xAI trained it on a wide range of agentic reinforcement-learning tasks, including kernel optimization, web development, and computer-aided design, and reports stronger first passes on visual and interactive projects plus more self-testing on long tasks. xAI says it matches GPT-5.6 Sol on the nine-benchmark Artificial Analysis Intelligence Index. (xAI)
What is new in GPT-5.6 Sol?
GPT-5.6 replaced the single-model-with-a-dial approach with three tiers: Sol (flagship), Terra (balanced), and Luna (cheapest). Sol adds a max setting for deeper reasoning and an ultra setting that coordinates four agents in parallel, plus Programmatic Tool Calling and multi-agent support in the Responses API. OpenAI cut Luna’s price 80% and Terra’s 20% on July 30, 2026. (OpenAI)
What is new in Claude Fable 5?
Claude Fable 5 is Anthropic’s first publicly available Mythos-class model, a tier above the Opus line, built for days-long autonomous work with sub-agents and self-checking. Anthropic calls it state-of-the-art on CursorBench and ships it with cybersecurity and biology safeguards that reroute flagged queries to an Opus model, so you are not charged Fable prices for rerouted requests. Using Fable 5 requires 30-day data retention for safety monitoring, part of Anthropic’s broader usage policy. (Anthropic)
Does Claude Fable 5 refuse some tasks the others answer?
Yes, and this is a real selection factor. Anthropic routes cybersecurity and biology queries to a less capable Opus model by default, and OpenAI noted that Fable 5 was excluded from its GeneBench Pro biology eval because it refuses the majority of those questions. For regulated defensive security or life-sciences work, that behavior may be a feature or a blocker depending on your use case. (Anthropic; OpenAI)
When were Grok 4.6, GPT-5.6 Sol, and Claude Fable 5 released?
Claude Fable 5 launched first, on June 9, 2026, with access restored July 1 after a brief pause. GPT-5.6 Sol reached general availability on July 9, 2026. Grok 4.6 is the newest, released August 12, 2026. (Anthropic; OpenAI; xAI)
Grok 4.6 vs GPT-5.6 Sol vs Claude Fable 5: which should you use?
Choose Grok 4.6 for the best price-to-capability ratio and output-heavy agent workloads, since it matches the others on intelligence indexes at a fraction of the token cost. Choose GPT-5.6 Sol for long-horizon agentic coding, terminal work, and parallel-agent runs via its ultra setting. Choose Claude Fable 5 for the hardest software-resolution tasks, where its SWE-Bench Pro lead and days-long autonomy justify the premium, provided its biology and cybersecurity refusals do not block your work. Treat all rankings here as provisional until an independent lab benchmarks all three on the same harness.
FAQs
Which is cheaper, Grok 4.6, GPT-5.6 Sol, or Claude Fable 5?+
Grok 4.6 is cheapest at $2 input and $6 output per million tokens. GPT-5.6 Sol is $5 and $30, and Claude Fable 5 is $10 and $50. Grok 4.6 is 80% cheaper on input and 88% cheaper on output than Fable 5, based on the makers’ August 2026 list prices (xAI; OpenAI; Anthropic).
Is Claude Fable 5 better than GPT-5.6 Sol?+
It depends on the task. On OpenAI’s own tests, Fable 5 leads SWE-Bench Pro (80% vs 64.6%) and the AA Intelligence Index v4.1 (59.9 vs 58.9), while Sol leads the Coding Agent Index, Terminal-Bench, and long-horizon Agents’ Last Exam. No single model wins across the board.
Does Grok 4.6 beat GPT-5.6 Sol?+
By xAI’s self-reported numbers, Grok 4.6 ties Sol on the Artificial Analysis Intelligence Index at 61 and leads on CursorBench, FrontierCode, and AA-Briefcase, but trails on DeepSWE and Terminal-Bench. These are xAI-run figures and were not independently confirmed at launch.
Why do the three models show different scores on the same benchmark?+
Each lab runs its own eval versions, harnesses, and reasoning settings. For example, xAI’s Terminal-Bench v3.0 scores are far lower than OpenAI’s Terminal-Bench 2.1 scores because they are different, harder tests. Compare within one maker’s table, not across tables.
Which model has the biggest context window?+
GPT-5.6 Sol and Claude Fable 5 both offer roughly a 1M-token context window, larger than Grok 4.6’s roughly 500K. Confirm the exact figures on each maker’s developer documentation before relying on them, as the makers’ marketing pages do not always state the number in plain text.
Is Grok 4.6 available yet?+
Yes. xAI released Grok 4.6 on August 12, 2026 as a significant improvement over Grok 4.5 at the same price. It is the newest of the three models, following GPT-5.6 Sol on July 9 and Claude Fable 5 on June 9.
