The full field, ranked, every model on the same yardsticks. Ordered left-to-right by overall intelligence rank. Numbers pulled July 12, 2026 from Artificial Analysis, SWE-bench (vals.ai), Terminal-Bench, GPQA & AA-Omniscience boards. Bold-green = best in that row.
| #1 overallFable 5Anthropic · our past experiment | #2 overallGPT-5.6OpenAI · "Sol" top variant | #5 overallOpus 4.8runs Archie & Foreman now | #8 overallGrok 4.5xAI · Max runs on this | #19 overallGemini 3.5Google · top public model | |
|---|---|---|---|---|---|
| Overall smarts & the basics | |||||
| Intelligence Index0–100, frontier is 55–60 | 60 | 59 | 56 | 54 | 50 |
| Context (memory) | 1M | 1M | 1M | 500K | 1M |
| Price · in / outper 1M tokens | $10 / $50 | $5 / $30 | $5 / $25 | $2 / $6 | $1.50 / $9 |
| Coding · Science · Doing-things | |||||
| CodingSWE-bench Verified | 95.0% | not published | 88.6% | 86.6% | 78.8% |
| Science reasoningGPQA Diamond | ~91–94%* | 94.6% | 93.6% | ~93% | 92.7% |
| Agentic tool-useTerminal-Bench 2.0 | 84.3% | 91.9% | 74.6% | 83.3% | 76.2% |
| Reliability & speed — the real-world stuff | |||||
| Hallucination ratelower = better; % confidently wrong | ~36% | not published† | 36% | 54% | not published |
| Output speedtokens/sec | 59 | 69 | 56 | 119 | 156 |
| Our own experience ◆ | refused benign work, benched | not run in-house | our rock-solid workhorse | Max runs clean for research | not run in-house |
*Anthropic retired GPQA (calls it saturated) so no official Fable 5 number; ~91–94% is third-party. †GPT-5.6 hallucination not yet published; the prior GPT-5.5 scored poorly here (guesses instead of abstaining). Google's flagship Gemini 3.1 Pro isn't public yet — 3.5 Flash is their top available model. Grok 4.5 launched 4 days ago; some of its scores are still firming up. ◆ = our own fleet tests, not internet numbers.
One 0–100 score that blends nine hard tests — real business-agent tasks, coding, PhD-level science, and factual knowledge — all run identically, one shot, no retries. The median model scores ~30; the frontier lives at 55–60. Roughly: a 60 gets ~60% of the hardest real tasks right on the first try.
500 real bugs pulled from big open-source projects. The model is dropped into the broken code, reads the bug report, and writes a fix — it never sees the tests. It only counts if its patch makes the broken tests pass and doesn't break anything else. 88% = it genuinely fixed 88% of 500 real bugs.
198 PhD-level science questions (biology, chemistry, physics) so hard that experts with Google barely pass and non-experts score near random guessing. Multiple choice, no tools. Top models now hit 93–95% — it's basically maxed out, so the leaders are all within 2 points of each other, effectively tied.
Can it actually do things, not just talk? It's handed a real command line and multi-step jobs — run code, fix a system, pull and analyze data — and has to chain many tool-calls, recover from errors, and finish. The closest test for "can this agent reliably operate a computer and fetch data." A model can ace single-question science yet drop 15 points here, because small mistakes pile up over a long task.
Good instinct. Grok is the cheapest, fastest, and most token-efficient, and strong at tool-calling — a real fit for high-volume web-fetching, which is exactly Max's lane. The catch: its hallucination rate doubled to 54%, so when it's pulling facts it'll sound certain about wrong things. So: yes, Max/Grok as the web-grabber — but only with grounding guardrails (below). If pure tool-reliability mattered more than cost, GPT-5.6 is actually #1 at agentic tool-use — worth knowing, but it's pricier and Grok is the better value for fetching.
You're right that wording fixes most of it. Highest-leverage first:
Feed it the source / live search — make it answer from a document you give it, not memory. Documented ~5× fewer made-up facts. Biggest single lever.
"Answer only from what I gave you — say 'I don't know' if it's not there." This one line is what makes Opus the best-calibrated model; you can bolt it onto Grok.
Demand citations / a self-check pass — have it quote its source and re-verify its own claims before finalizing.
Keep temperature low (~0.2–0.3) — minor, but helps.
Skip: generic "don't hallucinate" (does nothing), and long step-by-step reasoning for pure fact-lookup (actually makes it worse).
The benchmark winner isn't the one we run. Fable 5 tops the charts but flaked on us and got benched. Opus 4.8 stays the brain for anything that can't blink — Archie and Foreman. Grok/Max is the right cheap, fast engine for web-research and data-fetching, with the grounding guardrails. GPT-5.6 is the one to watch — it's the new #2 and the best pure tool-runner.
Sources: Artificial Analysis (Intelligence Index, speed, price, AA-Omniscience hallucination), SWE-bench Verified (vals.ai / benchlm.ai), GPQA Diamond & Terminal-Bench 2.0 (benchlm.ai), vendor cards. Prompting-mitigation figures: usewire.io, PromptHub, KeepMyPrompts, arXiv (CoVe / temperature studies). Harness choice can move coding scores ±5–15 pts; some Grok 4.5 & GPT-5.6 numbers are still propagating across independent evals.