GPT-5.6, Grok 4.5, Claude, and Muse Spark build the same 4 apps

In a single week in early July 2026, four major AI labs shipped or refreshed frontier models, and the team at TryAI.dev wasted no time putting all twelve resulting variants head-to-head in one of the most transparent coding challenges yet published: build four canonical apps from a cold prompt, five attempts each, raw outputs visible to anyone. The results complicate the tidy narratives every vendor has been selling — and they're essential reading for anyone deciding where to route API spend in the second half of 2026.

The Models and the Moment: A Crowded Week at the Frontier

The timing of this build-off matters as much as the results. Between July 8 and July 9, 2026, OpenAI, xAI, Meta, and Anthropic all had active frontier releases competing for developer attention — a convergence that turned the TryAI.dev challenge into an accidental industry snapshot.

OpenAI's GPT-5.6 launched on July 9 as a three-tier family, borrowing a naming convention from celestial bodies. Sol is the flagship, Terra is the balanced mid-range option, and Luna is the speed-and-cost play. All three share a February 2026 knowledge cutoff, a 1-million-token context window, and a 128,000-token maximum output per response. Pricing scales predictably: Sol costs $5/$30 per million input/output tokens, Terra $2.50/$15, and Luna $1/$6. New API capabilities — programmatic tool calling, multi-agent subagent spawning, and prompt cache breakpoints — ship across all three tiers, enabling developers to structure complex workflows without downgrading model capability. OpenAI has not yet published tier-specific SWE-Bench Pro scores for Terra or Luna.

xAI's Grok 4.5 arrived a day earlier, on July 8, at $2/$6 per million tokens with a 500,000-token context window — half that of the GPT-5.6 family. Its pitch is speed: roughly 80 tokens per second in independent tests, making it among the fastest frontier models available. This latency advantage makes it particularly attractive for real-time applications and high-throughput production systems.

Meta's Muse Spark 1.1 also landed on July 9, entering public preview on the Meta Model API. Priced at $1.25/$4.25 per million tokens, it is the most affordable closed frontier model in the comparison and carries a 1-million-token context window. Mark Zuckerberg described it on Threads as "a strong agentic and coding model at a very low price." Meta's broader product documentation elaborates on agentic coding as the model's primary focus: large-scale bug fixes, code migrations, and long-running tool-use and computer-use workflows that benefit from extended reasoning and iterative refinement.

Anthropic's Claude contributed two models: Claude Opus 4.8, an earlier release used as a reference point, and Claude Fable 5, a more recent and significantly more capable model priced at $10/$50 per million tokens — the most expensive in the field by a wide margin. This pricing premium reflects Anthropic's confidence in Fable 5's superior reasoning and code quality on high-stakes tasks.

Four open-weights models — Qwen 3.7 Plus, DeepSeek V4 Pro, Kimi K2.6, and GLM-5.2 — rounded out the twelve contestants, providing a cost baseline against which frontier pricing can be judged. These models allow teams to understand the economic trade-off between closed-model reliability and open-source flexibility.

The Four Apps: Deliberately Chosen to Stress Different Muscles

The TryAI.dev team selected their four tasks with care. Each probes a different dimension of a model's ability to translate a natural-language description into working, interactive code — without any intermediate guidance or error correction between attempts. This methodology mirrors real-world cold-start development scenarios.

1. Doom-style raycaster maze — WASD navigation, depth-shaded walls, collision detection. This is a novel, visually complex task with almost no single canonical implementation to copy from training data. It tests genuine compositional reasoning and spatial understanding.
2. 3D Rubik's Cube — scramble, solve, and animate a cube in three dimensions. Correctness is binary: the cube either solves cleanly or it doesn't. This task requires both mathematical precision and animation capability.
3. Calculator — operator precedence, proper styling, full arithmetic. Simple enough that every model should nail it; interesting as a floor test and as a style showcase. Well-represented in training data, making it a reliability baseline.
4. Conway's Game of Life — grid canvas with play, pause, step, randomize, and clear controls. Extremely well-represented in training data, making it a test of reliability rather than novelty. Thousands of reference implementations exist across GitHub, Stack Overflow, and technical blogs.

Each model received five cold attempts per task. The authors explicitly disclaim scientific objectivity — "This isn't objective" and "We are not handing down a scientific verdict" are direct statements in the piece — but they published every raw attempt for independent review, which is a higher standard of transparency than most AI benchmarks achieve. This commitment to showing raw outputs rather than summarized scores allows readers to form their own conclusions.

Task-by-Task: Where Each Model Shone and Where It Stumbled

Raycaster: GPT-5.6 Sol and Grok 4.5 Lead, Fable 5 Surprises With a Near-Miss

The Doom-style raycaster was the hardest task, and the frontier divide showed immediately. GPT-5.6 Sol went 5-for-5, producing playable, detailed renderers on every attempt — a consistency no other model matched on this task. GPT-5.6 Luna also went 5-for-5, though the authors noted its outputs were "not as good as GPT-5.5," suggesting the Luna tier trades some quality ceiling for cost and speed. This raises an important question about whether cheaper tiers represent true capability degradation or merely efficiency trade-offs.

Grok 4.5 also went 5-for-5, which is the result that demands the most attention: at $2/$6 per million tokens, producing consistently playable raycasters puts it in a genuinely compelling cost-performance position. The authors called it "great for price at 0.27¢" per reply on short prompts. This per-reply cost metric is more meaningful than raw token pricing for practitioners evaluating total cost of ownership.

Claude Opus 4.8 went 4-for-5, described as consistent but "dry" — technically adequate but lacking the visual polish of the top performers. Claude Fable 5 went 3-for-5 on the raycaster — a result that complicates its "most capable" billing on this specific task type, and a reminder that benchmark leadership does not always transfer to creative, generative, cold-start challenges. GPT-5.6 Terra also went 3-for-5.

The real shock on novelty tasks was Muse Spark 1.1, which went only 2-for-5. When it worked, the outputs were described as "surprisingly awesome," suggesting deep capability beneath the surface. But reliability was the problem. Meta's model is clearly still finding its footing on tasks that deviate sharply from well-trodden training territory. This is particularly relevant given Meta's preview status and its explicit focus on agentic rather than single-shot generation workflows.

Rubik's Cube: Claude Fable 5 Takes the Crown, Opus 4.8 and Luna Collapse

The Rubik's Cube task produced the biggest upsets of the entire comparison. Claude Fable 5 went a perfect 5-for-5, delivering clean solves with animations on every attempt — the standout individual performance of the entire build-off. This suggests that Anthropic's investment in spatial reasoning and mathematical precision on Fable 5 has paid off in concrete capability. GPT-5.6 Sol and Terra both went 4-for-5, maintaining their reliability even on a task that requires precise spatial and logical reasoning.

Grok 4.5 went 3-for-5 on the cube — a respectable but not dominant result, and a meaningful contrast to its 5-for-5 raycaster performance. The inconsistency on the cube suggests Grok 4.5's strengths are real but task-specific, reinforcing the message that no single model should be treated as a universal default.

Muse Spark 1.1 went 2-for-5 on the cube, matching its raycaster reliability score — a pattern suggesting the model's consistency issues span task categories, not just individual domains.

But Claude Opus 4.8 went 0-for-5. Zero clean solves across five attempts from a model that costs $5/$25 per million tokens is a result that should give any team running Opus 4.8 on code generation tasks serious pause. This result directly contradicts the narrative that Opus 4.8 is merely a cheaper Fable 5; the capability gap is categorical, not marginal. Equally striking: GPT-5.6 Luna also went 0-for-5 on the cube, with scrambling logic breaking the simulation on every attempt. This is the clearest evidence in the entire dataset that the Luna tier is not simply a cheaper Sol — it has a meaningfully lower capability ceiling on complex, novel problems. For teams considering Luna as a cost optimization, this result should trigger careful task-level testing before broad deployment.

Calculator: A Near-Sweep With Style Differences

The calculator was the great equalizer. Grok 4.5, Claude Opus 4.8, Claude Fable 5, GPT-5.6 Sol, GPT-5.6 Luna, and Muse Spark 1.1 all went 5-for-5. GPT-5.6 Terra went 4-for-5. On a task this well-represented in training data, every frontier model delivers near-perfect results. This validates the premise that training data saturation is one of the strongest predictors of model parity on code generation.

The differentiation shifted to aesthetics and implementation philosophy. Claude Fable 5 was named the authors' personal favorite for style, with clean, minimalist UI design. GPT-5.6 Sol was described as "overambitious with 3D styling" — it over-engineered the UI when a clean functional design was all that was asked for. This suggests that Sol's reasoning capability can lead to scope creep on simpler tasks. Muse Spark 1.1 produced results "on par with Grok," which is genuinely encouraging for a model only entering public preview. The lesson: on solved, low-novelty tasks, price becomes the primary variable, and architectural elegance becomes a tiebreaker.

Game of Life: Open Weights Compete on Economics

Conway's Game of Life is perhaps the most-implemented toy program in existence, making it the task most likely to be well-represented in every model's training data. On this task, the open-weights models — particularly Qwen 3.7 Plus and GLM-5.2 — demonstrated that abundant training examples can narrow the quality gap substantially. The TryAI.dev team did not publish formal pass/fail scores for this task, instead noting that cost and latency become the dominant selection criteria when correctness is table stakes; raw attempts for all models are available directly on the TryAI.dev site for independent review. This observation reinforces a critical pattern: frontier model pricing is justified primarily on novel, compositionally complex tasks; on canonical problems, open-weights alternatives offer compelling economics.

The Full Model Comparison: Specs, Scores, and Price

Model	Input ($/1M)	Output ($/1M)	Context	Raycaster	Rubik's Cube	Calculator	SWE-Bench Pro
GPT-5.6 Sol	$5.00	$30.00	1M	5/5	4/5	5/5	64.6%
GPT-5.6 Terra	$2.50	$15.00	1M	3/5	4/5	4/5	—
GPT-5.6 Luna	$1.00	$6.00	1M	5/5	0/5	5/5	—
Claude Fable 5	$10.00	$50.00	1M	3/5	5/5	5/5	80.0%
Claude Opus 4.8	$5.00	$25.00	1M	4/5	0/5	5/5	69.2%
Grok 4.5	$2.00	$6.00	500K	5/5	3/5	5/5	64.7%
Muse Spark 1.1	$1.25	$4.25	1M	2/5	2/5	5/5	61.5%

Table note: Dashes (—) indicate benchmarks not yet published by the model vendor as of July 2026. All build-off scores are from the TryAI.dev five-attempt cold-start methodology. SWE-Bench Pro figures are vendor-reported aggregates sourced from the BenchLM.ai leaderboard as of July 9, 2026.

The Price-Performance Reframe: Who Is Actually Cheap?

The raw benchmark scores tell one story; the price-adjusted story is more interesting for practitioners. Consider what these models cost at the task level: GPT-5.6 Luna's median reply latency was 1.0 seconds at roughly 0.001¢ per reply on short prompts. Terra came in at 1.5 seconds at the same token cost. Meanwhile, Grok 4.5 delivered its strong raycaster results at approximately 0.27¢ per reply. This per-reply framing is critical because it reveals that context matters: shorter prompts favor Luna and Grok 4.5 dramatically; longer prompts with higher output tokens favor upfront investment in Sol or Fable 5.

Claude Fable 5's $10/$50 pricing is a meaningful commitment. Its 80% SWE-Bench Pro score is the highest in this field by a substantial margin — roughly 15 full percentage points above GPT-5.6 Sol (64.6%) and Grok 4.5 (64.7%). For professional software engineering workflows where correctness is non-negotiable and volume is modest, that gap may justify the cost. On the other hand, GPT-5.6 Sol's Agents' Last Exam score of 53.6 has been publicly noted as exceeding Claude Fable 5's performance on that same long-horizon agentic benchmark — a counter-narrative Anthropic will need to address if agentic and tool-use workflows become the dominant deployment pattern. OpenAI has not yet published ALE scores for Terra or Luna, so direct comparison of those tiers with Fable 5 on that benchmark is not currently possible.

Why this matters: For the first time in the current generation, a developer can choose between five meaningfully distinct price-performance points — from Luna at $1/$6 to Fable 5 at $10/$50 — without falling off a cliff in capability. The right choice is now genuinely task-dependent, not brand-dependent. This fragmentation is healthy for the market and forces thoughtful procurement decisions rather than vendor lock-in reflexes.

Muse Spark 1.1: Meta's Most Credible Coding Model Yet

Meta's entry deserves its own section because it represents a qualitative shift in what the company is attempting. TechCrunch noted that Muse Spark 1.1 is specifically designed for large agentic workloads, bug fixes, and code migrations — enterprise-scale use cases that require sustained multi-step reasoning rather than single-prompt generation. On Threads, Mark Zuckerberg described it as "a strong agentic and coding model at a very low price." Meta's broader product documentation expands on this positioning, citing tool use and computer use as key capability areas — a notably different emphasis from OpenAI or Anthropic's messaging, which stresses breadth of capability across task types.

The build-off results paint a nuanced picture. On the raycaster — a novel, creative task — Muse Spark 1.1 went 2-for-5, which is genuinely below expectations for a model claiming frontier status. Yet the two successful outputs were described as "surprisingly awesome," suggesting the model has real capability that its reliability currently limits. On the Rubik's Cube, it went 2-for-5 — again, showing flashes of competence but falling short of consistent delivery. On the calculator, it matched Grok 4.5 exactly — a 5-for-5 clean sweep — further evidence that capability emerges unevenly across task categories.

On third-party benchmarks, Muse Spark 1.1 scores 61.5% on SWE-Bench Pro (behind Claude Fable 5's 80% and GPT-5.6 Sol's 64.6%, but competitive with Grok 4.5) and leads the field on MCP Atlas tool orchestration at 88.1%, which aligns with Meta's stated agentic focus. At $1.25/$4.25, it is cheaper than every other frontier model here by a meaningful margin. The TryAI.dev verdict — "pleasantly surprised" but "not something I would reach for just yet" — captures the trajectory accurately: not ready to lead on cold-start creative generation, but clearly not behind Anthropic or OpenAI by the significant gap that existed six months prior. For teams with existing integration into Meta's ecosystem and agentic workflows, Muse Spark 1.1 deserves serious pilot consideration.

The Open-Weights Wild Card: When Training Data Beats Frontier Pricing

The four open-weights models in the comparison — Qwen 3.7 Plus, DeepSeek V4 Pro, Kimi K2.6, and GLM-5.2 — were not competitive on the raycaster or Rubik's Cube. Novel, compositionally complex tasks still favor frontier closed models with more compute and, presumably, more RLHF investment in code generation. But Conway's Game of Life, which has thousands of reference implementations in every language across GitHub, Stack Overflow, and academic repositories, is a different story. Qwen 3.7 Plus, with a median latency of 2.1 seconds and approximately 0.001¢ per reply, matched frontier models on cost-normalized correctness for well-trodden tasks.

This is the pattern that will define open-weights adoption in 2026: wherever training data saturation is high and novelty is low, open-weights models eliminate the case for expensive frontier APIs. The gap persists — and is wide — on tasks that require genuine synthesis of rarely-combined concepts. The raycaster is a proxy for that entire category of problems. Teams considering whether to build closed-model or open-model pipelines should use this heuristic: closed models for novel synthesis, open-weights for canonical tasks at scale. Note that Google's Gemini 3.x was not included in the TryAI.dev build-off; a dedicated gpt gemini claude grok comparison — including Gemini's coding capabilities across all four task types — remains an open gap that future benchmarks are likely to fill as the Gemini family matures and competes more directly in these price bands.

GPT vs Grok vs Claude: What Each Camp Should Take Away

The gpt vs grok vs claude framing obscures as much as it reveals, but there are real strategic conclusions here. Comparing gpt gemini claude grok across the broader market — a comparison increasingly relevant as Gemini 3.x competes in the same price bands — the July 2026 picture looks like this:

• GPT-5.6 Sol is the most consistent top-tier performer across every task it attempted. It is the safest default for developers who need reliable results on novel, complex code generation and cannot afford to debug failures. Its 64.6% SWE-Bench Pro score and raycaster perfection make it the benchmark-setting generalist.
• Claude Fable 5 is the model for teams that need best-in-class SWE-Bench Pro performance (80%) and are willing to pay a 2–5× premium for it. Its Rubik's Cube perfection shows a model that reasons spatially and mathematically with high precision. It is the right choice for production software engineering where correctness on complex structured tasks is a hard requirement — though its 3/5 raycaster score is a genuine caveat for open-ended creative generation.
• Grok 4.5 is the value-tier story of the build-off. Matching GPT-5.6 Sol on the raycaster at less than half the output token cost makes it a legitimate secondary model or A/B candidate for any team currently defaulting to GPT-5.6 Terra. Its 3/5 Rubik's Cube result means it is not dominant across all task types, and its 500K context window is a meaningful limitation for large-file tasks. But for typical code generation workloads, the price advantage is compelling.
• Muse Spark 1.1 is for teams comfortable with a preview-tier model on agentic, long-context pipelines where its MCP Atlas score (88.1%) and Meta's stated tooling focus suggest it has real structural advantages — but not yet for cold-start creative code generation. As it matures, it may carve out a specific niche in enterprise migration and refactoring workflows.
• GPT-5.6 Luna is a trap on hard tasks. Its 0-for-5 Rubik's Cube failure alongside 5-for-5 raycaster performance means it requires careful task-routing: excellent for high-volume, well-defined work; unreliable for open-ended complexity. Teams should test Luna extensively on their specific workload before broad adoption.
• Claude Opus 4.8 should be deprecated in favor of Fable 5, Terra, or Grok 4.5. Its combined $5/$25 pricing and 0-for-5 Rubik's Cube failure make it difficult to articulate a use case where it outcompetes newer models at any price point.

Key Takeaways

• No single model wins every task. Claude Fable 5 leads SWE-Bench Pro (80%) and the Rubik's Cube; GPT-5.6 Sol and Grok 4.5 lead raycaster consistency; Grok 4.5 leads value-per-correct-output on complex generation. Task-specific dominance suggests that the future of AI development will involve model selection as a first-class API routing decision.
• Claude Fable 5's raycaster result is a genuine caveat. Its 3/5 on the most novel task — behind both GPT-5.6 Sol and Grok 4.5 — shows that SWE-Bench Pro leadership does not automatically translate to creative, cold-start generation. Teams should test Fable 5 on their actual workload before assuming it leads everywhere.
• Claude Opus 4.8 is functionally obsolete for hard coding tasks. A 0-for-5 Rubik's Cube result alongside a $5/$25 price tag makes it difficult to justify over GPT-5.6 Terra or Grok 4.5. Teams should begin migration planning immediately.
• GPT-5.6 Luna is not a budget Sol. Its capability ceiling is meaningfully lower on novel, compositionally complex problems — a critical distinction for teams tempted to cut costs by downgrading tiers. Per-token pricing savings should not drive a decision; per-task success rates should.
• Muse Spark 1.1 is Meta's most credible enterprise coding model yet, but its build-off reliability score (2/5 on both raycaster and Rubik's Cube) means it belongs in controlled pilot deployments, not production routing, for creative generation tasks. Its agentic benchmarks warrant investigation for teams with long-context or multi-step workflows.
• Grok 4.5 is the build-off's underrated story. Frontier-class raycaster and calculator results at $2/$6 per million tokens positions it as a genuine disruptor between the mid-tier and high-tier markets. Its 3/5 Rubik's Cube result prevents an unqualified sweep, but xAI's speed advantage (80 tokens/sec) adds practical value for latency-sensitive applications.
• Open-weights models close the gap on well-trodden tasks. For any code generation task with abundant training data, Qwen 3.7 Plus and GLM-5.2 eliminate the economic case for frontier APIs on that specific task category. Mixed workloads should adopt a hybrid model strategy.
• The "best AI" question is now permanently task-dependent. The spread between a $1/$6 Luna and a $10/$50 Fable 5 is too large for any team to ignore — and the results prove the cheaper option is not always inferior. Effective procurement now requires task categorization and model matching as a core infrastructure decision.
• Vendor benchmarks are increasingly insufficient. Published SWE-Bench Pro and Agents' Last Exam scores do not predict performance on task categories outside their training distribution. Orgs should run internal benchmarks on their actual workload before committing to a primary model.

---

The July 2026 build-off is a snapshot of a market in rapid, genuine fragmentation. Six months ago, the practical choice for serious code generation was a single flagship model per provider. Today, developers face a matrix of seven or more meaningfully differentiated options across price, context, speed, and task-specific capability. The next inflection point will likely be multi-agent routing: systems that automatically direct a prompt to Sol, Luna, Grok 4.5, or Muse Spark based on task complexity and cost budget. GPT-5.6's new programmatic tool calling and multi-agent subagent API features suggest OpenAI is already building the infrastructure for exactly that future — and if Muse Spark 1.1's agentic benchmarks hold up in production and the model's reliability improves, Meta may be closer behind than its build-off score implies.

For teams making AI infrastructure decisions in the second half of 2026, the message is clear: test your own workload, categorize your tasks by novelty and training-data saturation, and build a multi-model procurement strategy. The era of single-vendor dependency is functionally over. The era of routing and composition has begun.