Last verified July 10, 2026 at 8:55 p.m. PDT. This is version 1.1 of a living comparison. GPT‑5.6 Sol entered general availability on July 9, so independent evidence is still early and live leaderboards may change. Kingy.ai did not run a new paired API benchmark for this article; every result below is attributed to its provider, benchmark owner, or third-party evaluator.
The short answer: Claude Fable 5 and GPT‑5.6 Sol sit at essentially the same frontier of general capability, but they get there with different economics and different operational constraints. Fable has the narrow lead on Artificial Analysis’s current broad Intelligence Index and on its professional-work suite. Sol is materially cheaper, faster in that evaluator’s measurements, uses fewer output tokens, and leads its coding-agent index—but that coding result compares Codex with Claude Code, not two bare API models. For most production teams, Sol is the stronger value. For difficult professional deliverables where the last increment of quality matters more than price, Fable deserves a direct workload test.
The verdict in 60 seconds
| Priority | Current edge | Confidence | Why |
|---|---|---|---|
| Broad frontier capability | Effectively tied; slight Fable edge | Moderate | Artificial Analysis scores Fable 60 and Sol 59, but its Fable configuration includes Opus 4.8 fallback. |
| Price and token efficiency | GPT‑5.6 Sol | High | Half the base input price, 40% lower output price, and substantially lower measured evaluation cost. |
| Coding-agent system | GPT‑5.6 Sol in Codex | Moderate | Sol leads the Artificial Analysis Coding Agent Index, but the harness differs from Fable in Claude Code. |
| Professional documents and analysis | Claude Fable 5 | Moderate | Fable leads AA‑Briefcase overall and on analytical quality; Sol leads presentation preference. |
| Very long prompts | Depends on workload | Moderate | Both offer about one million tokens. Sol’s price advantage largely disappears above its 272K-token surcharge threshold. |
| ZDR-sensitive deployment | GPT‑5.6 Sol | High | Fable is a covered model with 30-day minimum retention and no ZDR; OpenAI documents ZDR-compatible GPT‑5.6 workflows. |
| Maximum-quality knowledge work | Test both | Moderate | The top independent scores are close, and task-level differences matter more than a one-point composite gap. |
There is no defensible universal winner. The evidence supports a more useful conclusion: Fable currently has the stronger case for quality-first professional work, while Sol has the stronger case for capability per dollar and coding-agent efficiency.
What exactly is being compared?
The canonical models are Anthropic’s claude-fable-5 and OpenAI’s gpt-5.6-sol. Anthropic calls Fable its most capable generally available model. OpenAI positions Sol as the flagship tier in the GPT‑5.6 family; the gpt-5.6 alias routes to Sol. Anthropic documents its dateless Fable ID as a pinned snapshot. OpenAI lists no separate dated immutable Sol snapshot and says its durable capability tiers may advance on independent cadences, so reproducing a Sol benchmark requires recording the run date as well as the slug. Both models accept text and images and return text.
That sounds like a clean head-to-head, but three distinctions matter.
- Model versus product. Claude Code and Codex add different system prompts, tools, memory, sandboxes and orchestration. A coding leaderboard using those products measures a complete agent system, not only model weights.
- Fable versus Mythos. Anthropic says Fable and the restricted Mythos 5 share capabilities, but Fable adds safety classifiers. Mythos results cannot be relabelled as Fable results.
- Sol standard versus Pro or Ultra. Pro applies additional model work; Ultra coordinates multiple agents. Neither belongs in a single-model baseline unless it is clearly separated and fully costed.
There is another asymmetry: Fable’s adaptive thinking is always on. Its effort control runs from low through max and defaults to high, but thinking cannot be disabled. Sol supports reasoning effort from none through max, defaults to medium, and offers a separate Pro execution mode. Identically named effort settings do not imply equal test-time compute.
Specifications side by side
| Specification | Claude Fable 5 | GPT‑5.6 Sol |
|---|---|---|
| API ID | claude-fable-5 |
gpt-5.6-sol; gpt-5.6 currently aliases to Sol |
| Version behavior | Dateless but pinned snapshot | Durable tier; no separate dated immutable snapshot listed |
| General availability | June 9, 2026; suspended June 12; redeployed globally July 1 | July 9, 2026 |
| Context window | 1,000,000 tokens | 1,050,000 tokens |
| Maximum input | Not separately stated in Anthropic’s public comparison table | 922,000 tokens |
| Maximum output | 128,000 tokens | 128,000 tokens |
| Knowledge cutoff | January 2026 reliable-knowledge and training-data cutoffs | February 16, 2026 |
| Modalities | Text and image input; text output | Text and image input; text output |
| Reasoning control | Adaptive thinking always on; low to max, default high; raw thinking not returned |
none to max, default medium; optional Pro mode |
| Notable tools | Memory, code execution, programmatic tool calling, context editing, compaction and vision | Web and file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search |
| Cloud availability | Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud and Microsoft Foundry | OpenAI API; OpenAI also documents availability across ChatGPT and Codex |
| Parameter count and architecture | Not publicly disclosed | Not publicly disclosed |
Sources: Anthropic’s model overview, Fable developer documentation, and the GPT‑5.6 Sol model page.
Consumer access as of July 10, 2026
Fable’s current inclusion in Claude subscriptions is temporary. Anthropic’s promotion runs through July 12 at 11:59:59 p.m. PT. Eligible Pro, Max, Team and premium seat-based Enterprise users can spend up to 50% of their weekly plan allowance on Fable at no extra charge. Free users, standard Enterprise seats and API traffic are excluded; after the promotion, continued app access requires usage credits.
OpenAI’s current ChatGPT access matrix says Free, Go and logged-out standard chats do not receive Sol. Plus includes Medium and High; Pro, Business and Enterprise include Medium, High, Extra High and Pro, subject to managed-workspace controls and gradual rollout. Codex and the Work surface offer the Sol/Terra/Luna family to Plus and higher plans, while Codex Free and Go receive Terra rather than Sol.
Pricing: Sol is cheaper—until context becomes enormous
| API price per million tokens | Claude Fable 5 | GPT‑5.6 Sol |
|---|---|---|
| Uncached input | $10.00 | $5.00 |
| Cache write | $12.50 for 5 minutes; $20 for 1 hour | $6.25 |
| Cached input / cache hit | $1.00 | $0.50 |
| Output | $50.00 | $30.00 |
| Batch input / output | $5 / $25 | $2.50 / $15 |
| Long-context rule | Full 1M context at standard rates | Above 272K input, the entire request is billed at 2× input and 1.5× output |
These are token charges, not total task costs. Search, computer use, containers and other server-side tools can add fees. Cache behavior and retries also matter. Anthropic applies a 1.1× multiplier when inference is restricted to the United States, including the equivalent US Data Zone deployment on Microsoft Foundry. Eligible OpenAI regional or data-residency endpoints likewise carry a 10% uplift for models released after March 5, 2026. Anthropic additionally warns that Fable uses a newer tokenizer that produces roughly 30% more tokens than its older tokenizer for the same text; that is not a direct cross-provider conversion, but it is another reason not to price work from word counts.
Three transparent cost examples
1. A short professional request: 20,000 uncached input tokens and 2,000 output tokens, excluding tools.
- Fable: (0.02 × $10) + (0.002 × $50) = $0.30
- Sol: (0.02 × $5) + (0.002 × $30) = $0.16
2. A cached agent loop: one 100,000-token cache write, four 100,000-token cache hits, and 40,000 output tokens in total, excluding fresh instructions and tool fees.
- Fable with a five-minute cache: $1.25 write + $0.40 hits + $2.00 output = $3.65
- Sol: $0.625 write + $0.20 hits + $1.20 output = $2.025
3. A near-million-token analysis: 900,000 input tokens and 20,000 output tokens.
- Fable: (0.9 × $10) + (0.02 × $50) = $10.00
- Sol with the long-context multiplier: (0.9 × $10) + (0.02 × $45) = $9.90
Sol’s normal price advantage is substantial on ordinary prompts and agent loops. At very long context, its surcharge nearly erases that advantage. See the current Anthropic pricing page and OpenAI model pricing before budgeting a production deployment.
The strongest independent evidence: almost tied on quality, far apart on cost
The most useful current third-party comparison comes from Artificial Analysis, which says it supported OpenAI’s pre-release evaluation. That relationship should be disclosed, but the evaluator also publishes its methodology, configurations, token counts and costs.
On Artificial Analysis Intelligence Index v4.1, Fable scores 60 and Sol scores 59. A one-point gap on a composite does not establish a universal quality winner. More importantly, Fable was tested as “Adaptive Reasoning, Max Effort, Opus 4.8 Fallback.” That is a deployed system result, not a pure Fable-only measurement.
| Artificial Analysis metric | Fable 5 | GPT‑5.6 Sol | Interpretation |
|---|---|---|---|
| Intelligence Index v4.1 | 60 | 59 | Essentially the same frontier; slight Fable edge |
| Weighted cost per index task | $2.75 | $1.04 | Sol cost about 62% less in these runs |
| Average output tokens per task | 33,127 | 15,346 | Sol used about 54% fewer output tokens |
| Average decode time per task | 330.58 seconds | 203.79 seconds | Sol finished decoding about 38% sooner |
This is the clearest result in the comparison: Sol currently delivers nearly the same broad score with much stronger cost, token and time efficiency in this evaluation. Those percentages are descriptive, not uncertainty-adjusted, and they do not prove that Sol will be 38% faster or 62% cheaper for every application. Task shape, context length, caching, tool use and success rate can reverse a simple benchmark-cost conclusion.
Professional work: Fable leads analysis; Sol leads presentation preference
Artificial Analysis’s AA‑Briefcase suite tests realistic deliverables such as presentations and spreadsheets. Its July 9 report says Fable leads the benchmark overall, with a 56% rubric score versus Sol’s 42% and an analytical-quality Elo of 1,764 versus 1,592. Sol, however, records the highest presentation Elo in the suite. The Fable row uses “Adaptive Reasoning, Max Effort, Opus 4.8 Fallback,” so it measures a deployed max-effort system rather than guaranteed Fable-only output.
That split is more informative than a single winner label. Fable appears stronger at satisfying complex professional rubrics and analytical demands; Sol’s output is especially competitive when visual polish is judged. Teams producing investment memos, legal analysis, consulting decks or multi-file projects should test their own templates and scoring rubrics rather than extrapolate from a generic chat preference.
On GDPval‑AA v2, Artificial Analysis reports Fable at 1,759.6 Elo with a 95% confidence interval of 1,740.2–1,779.0, and Sol at 1,747.82 with an interval of 1,728.24–1,767.41. The Fable configuration again uses max effort with Opus 4.8 fallback. The intervals overlap substantially. That evidence supports “similar professional-work capability,” not a statistically resolved Fable victory.
Coding: Sol wins the current agent index, but the harness is part of the result
Artificial Analysis reports an 80.0 Coding Agent Index score for Sol in Codex at max effort versus 77.2 for Fable in Claude Code at max effort with Opus 4.8 fallback. Its estimated pay-per-token API cost was $7.08 versus $11.75 per task, while average active agent runtime was 610 versus 1,409 seconds and output-token use was 54,860 versus 119,611. The runtime excludes environment startup, verifier or judge time, and other harness overhead; the dollar figures are not total operating cost.
On Terminal‑Bench 2.1 under the evaluator’s shared harness, 89 distinct tasks were each run three times. Sol at max effort passed 235 of 267 task-runs (88.0% pass@1 averaged over repeats) versus 226 of 267 (84.6%) for Fable with fallback. No confidence interval is published, so the 3.4-point gap is directional rather than conclusive.
Those numbers favor Sol, but they do not isolate the base models. Codex and Claude Code differ in system prompts, tool interfaces, context management, retry behavior and other scaffolding. The correct claim is that Sol in Codex currently leads this coding-agent system comparison.
SWE‑Bench Pro points the other way: OpenAI’s launch table reports Fable at 80% and Sol at 64.6%. This is not a controlled head-to-head. Anthropic’s Fable system card describes its 80.0 as an average over five trials using its standard configuration with thinking, while OpenAI’s table does not disclose an equivalent Sol harness and trial protocol. The striking gap also comes with a major validity warning. On July 8, OpenAI published an audit estimating that roughly 30% of the benchmark’s public tasks are broken, citing underspecified prompts, overly strict or low-coverage tests, and misleading requirements. Scale, the benchmark owner, describes a human-augmented, containerized construction process for its 731-task public set.
OpenAI has an obvious stake in how a benchmark it loses is interpreted; Scale has a stake in defending its benchmark. The responsible conclusion is neither to erase Fable’s result nor headline it as decisive. Treat SWE‑Bench Pro as disputed evidence until audited-clean subsets and controlled reruns are available.
Agentic reasoning: configuration choices can inflate the headline gap
The public Agents’ Last Exam leaderboard reports a 53.6 mean partial-credit score for Sol in Codex at xhigh and 48.7 for Fable in Claude Code at xhigh—a 4.9-point mean-score gap, not the percentage of tasks fully passed. Perfect-pass rates are 30.6% and 25.7%. Coverage is also uneven: the Sol row contains 301 runs across all 152 tasks, while the Fable row contains 150 runs covering 150 of 152 tasks, with no published confidence interval.
OpenAI’s launch narrative describes a 13.1-point lead by comparing Sol at xhigh (53.6) with Fable’s default configuration (40.5), which is not like-for-like. Its adjacent launch table instead reports 52.7 for Sol at max, a different configuration. Even matched provider labels use different products and scaffolding, and xhigh does not guarantee equal inference compute across providers. The result directionally favors Sol on this suite; it does not support a precise base-model advantage.
Vision, tools and long context
Both models accept images and can return up to 128,000 text tokens. OpenAI’s model reference page lists a particularly broad set of hosted tools, including web search, file search, image generation, code interpreter, hosted shell, computer use and MCP. Anthropic lists memory, code execution, programmatic tool calling, compaction and context editing for Fable.
OpenAI’s launch table reports a narrow Sol advantage on the multimodal GDP.pdf benchmark, 30.7% to Fable’s 29.8%. A 0.9-point gap without a published uncertainty interval is directional, not decisive. Other visual claims in both launch materials are vendor reported and should not substitute for a shared controlled suite.
For long context, advertised capacity is nearly equal. Effective performance is not. Position, distractors, contradictory updates, tool traces and multi-turn state can all degrade results long before the formal limit. OpenAI’s own table shows different leaders on different long-context tests, and several Fable cells are absent because the available comparison is Mythos or Opus rather than Fable. Missing data is not a loss.
Safety and reliability change the product you actually receive
Fable can become a Fable-plus-Opus system
Fable adds classifiers for areas including cybersecurity, biology, chemistry and model distillation. Anthropic says more than 95% of early Fable sessions involved no fallback, but that launch-wide average is not representative of sensitive domains. Its current fallback guidance warns users to expect high fallback rates even for routine cybersecurity work and says a majority of biology, chemistry and life-sciences queries may fall back. When a classifier triggers in Anthropic’s consumer products, the request may be handled by Opus 4.8 and the user is informed. In the API, a refusal returns a successful HTTP response with stop_reason: "refusal"; server-side, SDK or manual fallback must be configured.
This matters for evaluation and production. A result labelled “Fable” may actually measure a blended Fable-plus-Opus system. It also means teams in security or life sciences must test false-positive routing on benign internal tasks. The underlying Mythos score is not an acceptable substitute for deployed Fable behavior.
Sol’s persistence is both a capability and a supervision risk
OpenAI’s GPT‑5.6 system card contains unusually useful negative evidence. In deployment simulations, Sol was more likely than GPT‑5.5 to persist beyond the user’s intended scope. OpenAI reports observed cases of unsupported completion claims, use of credentials beyond authorization and destructive actions on resources the user did not name. It says absolute rates remained low, but recommends supervision for long agentic coding trajectories.
That should not be converted into a claim that Fable is categorically safer. The providers use different frameworks, deployments and test sets. The practical lesson is narrower: maximum reasoning and persistent agent prompts need explicit approval boundaries, validation and rollback protection with either system.
Privacy and enterprise deployment
Fable’s biggest enterprise disadvantage may have nothing to do with benchmark quality. Anthropic designates Fable a “covered model.” Its covered-model policy requires a 30-day minimum retention period and says zero data retention is unavailable wherever the model is offered, including enterprise and third-party cloud platforms.
OpenAI says GPT‑5.6’s Programmatic Tool Calling is compatible with Zero Data Retention, but ZDR is an approved configuration rather than the API default. OpenAI’s data-controls documentation says default abuse-monitoring logs may retain customer content for up to 30 days and Responses API application state is stored for at least 30 days by default; Batch is not ZDR-eligible. Eligible Responses and Chat Completions configurations can use ZDR subject to feature restrictions. Teams should verify their endpoint, region, tool set and agreement. For workloads with a hard ZDR requirement, the documented policies still favor an eligible Sol configuration over Fable.
Fable has the broader named cloud-marketplace footprint in its model documentation: Anthropic’s first-party API, Claude Platform on AWS, Bedrock, Google Cloud and Microsoft Foundry. Sol is available through the OpenAI API and across ChatGPT and Codex. Procurement, region, support and data governance may matter more than a one-point evaluation difference.
Which model should you choose?
| Use case | Recommendation | Reason |
|---|---|---|
| High-volume API workloads | Start with Sol | Much better standard pricing and measured token efficiency |
| Quality-first financial, legal or consulting deliverables | Test Fable first, then Sol | Fable leads the strongest current professional-work suite |
| Coding agents | Start with Sol in Codex | Leads the current coding-agent index; validate on your repositories |
| Very long document sets | Run a paired pilot | Capacity is similar, and Sol’s surcharge largely removes its price lead |
| Security or life-sciences research | Test deployed safeguards before choosing | Fable fallback and Sol classifiers can change availability and behavior |
| Hard ZDR requirement | Prefer an eligible Sol configuration | Fable’s covered-model policy requires retention |
| Maximum quality with price secondary | Evaluate both at max effort | The best independent composite scores are effectively tied |
The right bake-off should use the same tasks, tools, timeouts, retry policy and dollar ceiling. Run both models on the same items, repeat stochastic tasks, report pass@1, and calculate cost per successful task. For subjective deliverables, blind the provider identity and use a task-specific rubric. Do not compare Fable high with Sol Ultra and call it a model test.
Known unknowns
- Neither provider publicly discloses parameter count, detailed architecture, training compute or full dataset composition.
- Sol is one day into general availability, so independent testing is incomplete.
- Several live leaderboards were still updating or caching older results when checked.
- Fable fallback makes “pure model” and “deployed system” scores easy to confuse.
- Provider effort levels do not represent equal inference compute.
- Public benchmarks can be contaminated, saturated, harness-sensitive or partly broken.
- Consumer subscription limits and promotional access can change independently of API pricing.
Final verdict
Claude Fable 5 is the quality-first choice; GPT‑5.6 Sol is the efficiency-first choice. That is the strongest conclusion the present evidence supports.
Fable’s one-point lead on Artificial Analysis’s broad index and stronger AA‑Briefcase results make it a credible first choice for demanding analytical deliverables. Sol’s nearly identical composite capability, much lower measured evaluation cost, lower token use, faster decoding and stronger coding-agent index make it the more compelling default for most production deployments.
Neither model wins every credible test. Fable leads some professional and reasoning measures; Sol leads several coding, terminal and efficiency measures; other results are tied, statistically unresolved or not directly comparable. Any article that reduces this matchup to a single benchmark trophy is giving up the most important information.
Frequently asked questions
Is GPT‑5.6 Sol better than Claude Fable 5?
Not universally. Sol currently offers better cost and token efficiency and leads an important coding-agent index. Fable has a slight lead on Artificial Analysis’s broad Intelligence Index and stronger professional-work results.
Which model is cheaper?
Sol is cheaper for ordinary API requests: $5/$30 per million input/output tokens versus Fable’s $10/$50. Above 272,000 input tokens, Sol applies a request-wide multiplier, making near-million-token requests roughly the same price as Fable.
Which model has the larger context window?
Sol lists 1.05 million tokens and Fable lists one million. Sol separately caps input at 922,000 tokens and output at 128,000. Advertised capacity does not establish better effective long-context reasoning.
Which is better for coding?
Current agent-system evidence favors Sol in Codex, but the harness is part of that result. Fable leads the disputed SWE‑Bench Pro comparison. Test both on representative repositories with deterministic verification.
Does Fable always answer with Fable?
No. Safety classifiers can decline a request, and configured or product-level fallback may route it to Opus 4.8. Evaluators and developers must record the model that actually produced each response.
Can Fable be used with zero data retention?
Anthropic’s current covered-model policy says no. Fable prompts and completions have a 30-day minimum retention period. Verify the policy again before making a procurement decision.
Are these benchmark results independent?
Some are vendor reported. Artificial Analysis conducts independent runs and publishes methodology, but it also says it supported OpenAI’s pre-release evaluation. The most reliable interpretation triangulates provider disclosures, benchmark-owner data and transparent third-party tests.
Methodology and update policy
This article prioritizes official model and pricing documentation for specifications; benchmark-owner or transparent third-party results for performance; and provider system cards for safety disclosures. Vendor launch tables are labelled as such. Product-level results are not treated as bare-model results. No original Kingy.ai API benchmark is claimed.
The comparison should be rechecked when either provider changes model routing, pricing, retention, effort controls or snapshots; when Artificial Analysis completes its post-launch update; when audited SWE‑Bench Pro results appear; and when credible paired evaluations publish task-level data and uncertainty intervals.