Every few weeks someone asks “which model should I use?” and the answer changes before the reply is posted. Here’s a systematic comparison of every frontier API model from Anthropic and OpenAI as of June 25, 2026, evaluated on output throughput, latency, cost efficiency, and intelligence — with actual numbers instead of vibes.
The methodology borrows from academic Systematization of Knowledge (SoK) papers: decompose the problem into orthogonal dimensions, define objective properties, fill a matrix, and let the patterns speak for themselves. All speed and intelligence data comes from Artificial Analysis (independent, 72-hour median measurements) unless otherwise noted. Pricing from official docs.
The landscape
Anthropic currently ships a clean four-tier ladder: Haiku (speed) → Sonnet (balanced) → Opus (frontier) → Fable (autonomous). Each roughly doubles in price and steps up in intelligence.
OpenAI has converged on effort levels within a single model family. GPT-5.5 at low/medium/high/xhigh gives a continuous quality-speed dial rather than discrete tiers. GPT-4o persists as a legacy speed workhorse.
Throughput & latency
This is where model selection gets interesting. Two models dominate the speed tier:
| Model | Output t/s | TTFT | Input / Output per M |
|---|---|---|---|
| Haiku 4.5 | ~92 | 0.78s | 5 |
| GPT-4o | ~117 | 0.96s | 10 |
| Sonnet 4.6 | 40–60 | ~2s (non-reasoning) | 15 |
| Opus 4.8 | ~61 | ~31s (reasoning) | 25 |
| Opus 4.8 Fast | ~150 est. | ~31s | 50 |
| Fable 5 | no data yet | unknown | 50 |
| GPT-5 (high) | ~82–98 | ~70s | 10 |
| GPT-5.5 (medium) | ~59 | ~12s | 30 |
| GPT-5.5 (high) | ~68 | ~28s | 30 |
Haiku 4.5 and GPT-4o both achieve sub-second TTFT and are suitable for real-time interactive use. The critical difference is intelligence: Haiku scores 24 on the Artificial Analysis Intelligence Index versus GPT-4o’s 17. Haiku is both faster-feeling and smarter.
The mid-tier (50–70 t/s) clusters Sonnet 4.6, Opus 4.8, GPT-5 (high), and GPT-5.5 variants. Sonnet 4.6 stands out for non-reasoning workloads with low TTFT. Opus 4.8 and GPT-5.5 pay a steep TTFT penalty from reasoning overhead — 30 seconds and 12–89 seconds respectively. This makes them unsuitable for real-time chat without paying for Fast Mode.
Opus 4.8 Fast Mode is unique in the market: same model at 2.5x speed for 2x price. At ~150 t/s effective throughput with frontier intelligence, it’s the only option that doesn’t force a speed-intelligence tradeoff. Previous Opus fast modes cost roughly 150 per million tokens — the drop to 50 is structural.
Fable 5 launched June 9, 2026 and has no published TPS benchmarks yet. Given its Mythos-class architecture and long-horizon agentic focus, expect speed comparable to or slower than Opus 4.8.
Intelligence & task fitness
The intelligence landscape splits into three tiers:
Tier 1 (Intelligence Index 55+): Fable 5 (60), Opus 4.8 (56), GPT-5.5 xhigh (55). These are the reasoning-heavy frontier models. Fable 5 leads agentic coding with 80.3% on SWE-bench Pro. Opus 4.8 leads computer use (83.4% OSWorld) and real-world work tasks (1890 Elo on GDPval-AA). GPT-5.5 leads Terminal-Bench (78.2% vs Opus 4.8’s 74.6%).
Tier 2 (Intelligence Index 45–55): Sonnet 4.6 (51), Opus 4.6 (53), GPT-5.5 high (53). These deliver 90%+ of Tier 1 quality for most tasks. Sonnet 4.6 is the standout value play at 15. Developers preferred it over Opus 4.5 (the previous flagship) in 59% of Claude Code sessions.
Tier 3 (Intelligence Index < 30): Haiku 4.5 (24), GPT-4o (17). Fast execution models. Haiku 4.5 matches the coding quality of the original Claude Sonnet 4, which was frontier six months prior.
A notable asymmetry: Anthropic dominates agentic coding and computer use. OpenAI leads terminal/CLI automation (Terminal-Bench). For pipelines that drive browsers or navigate codebases, Opus 4.8 and Fable 5 are ahead. For shell-centric workflows, GPT-5.5 has a genuine edge.
The cost-efficiency trap
Raw per-token pricing is misleading without accounting for token efficiency. Three examples that illustrate the gap between sticker price and real cost:
Sonnet 4.6 (25 output): Sonnet uses ~28% more output tokens than Opus 4.8 on Intelligence Index evaluations (74M vs 58M). On reasoning-heavy tasks, the effective cost gap narrows to ~15–20% rather than the nominal 40%.
Opus 4.8 (30 output): Opus 4.8 uses ~35% fewer tokens than Opus 4.7 and completes agentic tasks with higher pass rates, meaning fewer retries. GPT-5.5 may cost more per successful task on coding workloads despite nominally lower output pricing.
Haiku 4.5 (10 output): Haiku is both cheaper and smarter. GPT-4o’s only advantage is raw speed (117 vs 92 t/s), which rarely matters in practice.
Anthropic’s caching story is strong: cached input tokens cost ~$0.50/M on Opus-class models, cheaper than Haiku’s standard input rate. For workloads with repeated system prompts (agentic loops, coding assistants), caching dramatically compresses effective input cost across all Claude models.
The Pareto frontier
Plot intelligence against output speed and the efficient frontier becomes clear:
- Haiku 4.5 (92 t/s, II 24) — Pareto-optimal for speed-weighted workloads
- Sonnet 4.6 (53 t/s, II 51) — Pareto-optimal for balanced workloads
- Opus 4.8 (61 t/s, II 56) — Surprisingly close to Sonnet on speed with higher intelligence
- Fable 5 (? t/s, II 60) — Likely Pareto-optimal for intelligence-weighted workloads
GPT-4o (117 t/s, II 17) is not on the Pareto frontier — Haiku 4.5 dominates it with higher intelligence at comparable speed and lower cost. GPT-5.5 high (68 t/s, II 53) is close to Opus 4.8 but Opus leads on most coding benchmarks at lower output pricing (30).
Forced tradeoffs
Three tradeoffs that no amount of model selection can escape:
Speed vs TTFT on reasoning models. Opus 4.8 and GPT-5.5 both achieve decent output t/s once generating, but their reasoning overhead creates 12–89 second TTFT. This makes them unsuitable for real-time chat without the Fast Mode premium.
Token efficiency vs per-token price. Sonnet 4.6’s low per-token price is partially offset by verbose output. Opus 4.8’s higher price is offset by token efficiency. The “cheaper model” is frequently not cheaper per task.
Context window vs retrieval quality. GPT-5.5’s 922K window is the largest, but published tests show GPT-4o degrades past 64K. Claude models generally maintain better long-context retrieval fidelity, though independent benchmarks at 922K scale are sparse.
What to actually use
For an indie developer running agentic coding pipelines on a budget:
Default model: Sonnet 4.6. Best intelligence-per-dollar for daily coding. The workhorse.
Sub-agent / router / high-volume: Haiku 4.5. Classifier, code review triage, simple completions. 6x cheaper than Sonnet on output with near-instant response.
Hard problems / codebase refactoring: Opus 4.8. Reserve for tasks where Sonnet fails or correctness is critical. The 35% token efficiency improvement over 4.7 makes it surprisingly affordable for targeted use.
Long-horizon autonomous work: Fable 5. Only when the task genuinely warrants multi-hour autonomous execution. At 50, use sparingly with budget caps.
Terminal-heavy CLI workflows: GPT-5.5 (medium). If your pipeline is shell-centric and Terminal-Bench performance matters, GPT-5.5 has a genuine edge here.
Skip entirely: GPT-4o, Opus 4.6. GPT-4o is dominated by Haiku 4.5 on every axis except raw t/s. Opus 4.6 is superseded by 4.8 at the same price.
What we still don’t know
Fable 5 speed data doesn’t exist yet — the model is 16 days old and Artificial Analysis hasn’t benchmarked it. Token efficiency standardization across different tokenizers remains an unsolved problem. Reasoning token costs are opaque from both providers. Cross-provider speed variance can reach 30–50% on the same model. And any static recommendation has a half-life of weeks — Opus went from 4.7 to 4.8 in 41 days.
Bias disclaimer
This comparison was produced by Claude (an Anthropic model). All speed and intelligence data comes from independent third-party benchmarks, but interpretation may unconsciously favor Anthropic’s framing. Gemini 3.1 Pro, DeepSeek V4, Llama 4, and Qwen3 are excluded — a complete comparison would include them.