Tier 2: The Specialists and Speed Demons

Claude Opus 4.8 (Anthropic)

Before Fable 5's arrival, Opus 4.8 was the undisputed king. Released in late May 2026, it's now the powerful second-in-command at Anthropic and remains a top-tier model in its own right .

• Overall & Agentic: It was the first model to lead the Artificial Analysis Intelligence Index with a score of 61.4 and tops the GDPval-AA leaderboard at 1890 Elo, before being surpassed by Fable 5 .
• Coding: It scores 69.2% on SWE-bench Pro and 88.6% on SWE-bench Verified .
• Math: It demonstrated a historic 27-point leap on the USAMO 2026 competitor math benchmark, jumping from 69.3% to 96.7% in 41 days .
• Honesty: The model has the lowest rate of factual hallucination among its peers, primarily by abstaining when uncertain rather than guessing .

GPT-5.5 (OpenAI)

GPT-5.5, released in late April 2026, was the first model in this generation to retake the overall performance lead for OpenAI, but with a more specialized skill set .

• Mathematical Reasoning: It is the stand-out math model, scoring 51.7% on FrontierMath Tier 1-3 and 35.4% on the impossibly hard Tier 4—a clear lead over the field .
• Autonomous CLI: It also leads on Terminal-Bench 2.0, scoring 82.7% for complex command-line tasks, compared to Claude Opus 4.7's 69.4% .
• Long Context: GPT-5.5 made a 37-point leap in long-context retrieval (MRCR v2), scoring 74.0% at a 1M-token context, leaving its predecessor's 36.6% in the dust .
• Caveat on Coding: While a strong all-rounder, its SWE-Bench Pro score of 58.6% now significantly lags behind Anthropic's models .

Gemini 3.5 Flash (Google)

Debuted at Google I/O 2026, this model is a 'Flash' tier offering, meaning it's designed for speed and cost-efficiency without sacrificing the ability to punch above its weight .

• Speed and Value: It is reportedly 4x faster than Gemini 3.1 Pro for agent and coding tasks, with competitive pricing .
• Agentic & Tool Use: It scores a strong 83.6% on MCP Atlas, a multi-step tool-use benchmark .
• Coding & Vision: On SWE-Bench Pro, its score of 55.1% slightly surpasses its larger predecessor, Gemini 3.1 Pro, and it scores 84.2% on the CharXiv visual reasoning benchmark .
• Reality Check on Coding: Independent Artificial Analysis data suggests its coding scores are significantly weaker than Google's own comparisons suggest, so its real-world utility for developers is still being validated .

Grok 4.3 (xAI)

Launched quietly at the end of April 2026, Grok 4.3 is a cost-efficiency and speed play that excels in niche areas .

• Overall Standing: It scores 53 on the Artificial Analysis Intelligence Index. This places it above mid-tier models like Claude Sonnet 4.6 but well below the top-tier GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro .
• King of Instruction Following: It ranks #1 on IFBench with an 81% score, beating GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.7 .
• Agentic Boom: Grok 4.3 shows a massive +321 Elo point jump on the GDPval-AA agentic tasks benchmark over its predecessor, reaching a score of 1500 . It also scores a field-leading 98% on the τ²-Bench Telecom benchmark .
• Speed Demon: Early reports list its speed at 159 tokens per second, with other benchmarks showing speeds between 78.4 and 196 tokens/s .
• Pricing & Limits: Priced aggressively at an estimated $1.25/$2.50 per million tokens (input/output), it is a strong value play for specific domains but is not a general intelligence competitor .

The Open-Source Contender: DeepSeek V4 Pro (DeepSeek AI)

Released in April 2026, DeepSeek V4 Pro is the strongest open-source challenger, offering incredible coding value but with a critical independent caveat.

• Coding Value Leader: It scores an impressive 80.6% on SWE-bench Verified and 93.5% on LiveCodeBench, putting it within a hair's breadth of Claude Opus 4.6 on coding at a fraction of the price (estimated ~$3.48/M output tokens) .
• The CAISI Verdict: An independent evaluation by the U.S. government's CAISI found that DeepSeek V4's capabilities lagged the frontier by about 8 months. While DeepSeek's own data compared it to GPT-5.4, CAISI pegs its performance closer to GPT-5 from August 2025 .
• Overall Ranking: It lags on general knowledge and complex reasoning tasks compared to proprietary giants, ranking #39 out of 123 models on one broader leaderboard .

Safety & Independent Evaluations

Independent, standardized safety testing remains rare and often unpublished.

• DeepSeek V4 Pro (CAISI/NIST): The U.S. government's evaluation remains the most notable safety signal, concluding its capabilities were about 8 months behind the frontier .
• GPT-5.5 (AI Security Institute): The UK's AI Security Institute reported GPT-5.5 had a 71.4% pass rate on its expert-level cyber tasks, noting it "may be the strongest model we have tested" on that measure .

How to Think About These Results

It's tempting to crown a single winner, but the data paints a more practical picture:

• For pure coding ability and agentic tasks, Claude Fable 5 is in a class of its own.
• For heavy mathematical reasoning and autonomous command-line work, GPT-5.5 is the tool of choice.
• For a balanced, cost-effective model that's fast and reliable, Gemini 3.5 Flash is a compelling option, though you should validate its coding claims for your specific use case.
• If you need a specialist for legal reasoning, finance, or instruction-following at a low price, Grok 4.3 is a surprisingly potent choice.
• If you're building on a budget and need frontier-grade coding with an open-source license, DeepSeek V4 Pro is unmatched in value, but its real-world sophistication may not match the very best.

Always remember that benchmark conditions vary (prompting, tool use, reasoning effort) and aren't a perfect predictor of your specific application.

Tier 2: The Specialists and Speed Demons

Claude Opus 4.8 (Anthropic)

Before Fable 5's arrival, Opus 4.8 was the undisputed king. Released in late May 2026, it's now the powerful second-in-command at Anthropic and remains a top-tier model in its own right .

• Overall & Agentic: It was the first model to lead the Artificial Analysis Intelligence Index with a score of 61.4 and tops the GDPval-AA leaderboard at 1890 Elo, before being surpassed by Fable 5 .
• Coding: It scores 69.2% on SWE-bench Pro and 88.6% on SWE-bench Verified .
• Math: It demonstrated a historic 27-point leap on the USAMO 2026 competitor math benchmark, jumping from 69.3% to 96.7% in 41 days .
• Honesty: The model has the lowest rate of factual hallucination among its peers, primarily by abstaining when uncertain rather than guessing .

GPT-5.5 (OpenAI)

GPT-5.5, released in late April 2026, was the first model in this generation to retake the overall performance lead for OpenAI, but with a more specialized skill set .

• Mathematical Reasoning: It is the stand-out math model, scoring 51.7% on FrontierMath Tier 1-3 and 35.4% on the impossibly hard Tier 4—a clear lead over the field .
• Autonomous CLI: It also leads on Terminal-Bench 2.0, scoring 82.7% for complex command-line tasks, compared to Claude Opus 4.7's 69.4% .
• Long Context: GPT-5.5 made a 37-point leap in long-context retrieval (MRCR v2), scoring 74.0% at a 1M-token context, leaving its predecessor's 36.6% in the dust .
• Caveat on Coding: While a strong all-rounder, its SWE-Bench Pro score of 58.6% now significantly lags behind Anthropic's models .

Gemini 3.5 Flash (Google)

Debuted at Google I/O 2026, this model is a 'Flash' tier offering, meaning it's designed for speed and cost-efficiency without sacrificing the ability to punch above its weight .

• Speed and Value: It is reportedly 4x faster than Gemini 3.1 Pro for agent and coding tasks, with competitive pricing .
• Agentic & Tool Use: It scores a strong 83.6% on MCP Atlas, a multi-step tool-use benchmark .
• Coding & Vision: On SWE-Bench Pro, its score of 55.1% slightly surpasses its larger predecessor, Gemini 3.1 Pro, and it scores 84.2% on the CharXiv visual reasoning benchmark .
• Reality Check on Coding: Independent Artificial Analysis data suggests its coding scores are significantly weaker than Google's own comparisons suggest, so its real-world utility for developers is still being validated .

Grok 4.3 (xAI)

Launched quietly at the end of April 2026, Grok 4.3 is a cost-efficiency and speed play that excels in niche areas .

• Overall Standing: It scores 53 on the Artificial Analysis Intelligence Index. This places it above mid-tier models like Claude Sonnet 4.6 but well below the top-tier GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro .
• King of Instruction Following: It ranks #1 on IFBench with an 81% score, beating GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.7 .
• Agentic Boom: Grok 4.3 shows a massive +321 Elo point jump on the GDPval-AA agentic tasks benchmark over its predecessor, reaching a score of 1500 . It also scores a field-leading 98% on the τ²-Bench Telecom benchmark .
• Speed Demon: Early reports list its speed at 159 tokens per second, with other benchmarks showing speeds between 78.4 and 196 tokens/s .
• Pricing & Limits: Priced aggressively at an estimated $1.25/$2.50 per million tokens (input/output), it is a strong value play for specific domains but is not a general intelligence competitor .

The Open-Source Contender: DeepSeek V4 Pro (DeepSeek AI)

Released in April 2026, DeepSeek V4 Pro is the strongest open-source challenger, offering incredible coding value but with a critical independent caveat.

• Coding Value Leader: It scores an impressive 80.6% on SWE-bench Verified and 93.5% on LiveCodeBench, putting it within a hair's breadth of Claude Opus 4.6 on coding at a fraction of the price (estimated ~$3.48/M output tokens) .
• The CAISI Verdict: An independent evaluation by the U.S. government's CAISI found that DeepSeek V4's capabilities lagged the frontier by about 8 months. While DeepSeek's own data compared it to GPT-5.4, CAISI pegs its performance closer to GPT-5 from August 2025 .
• Overall Ranking: It lags on general knowledge and complex reasoning tasks compared to proprietary giants, ranking #39 out of 123 models on one broader leaderboard .

Safety & Independent Evaluations

Independent, standardized safety testing remains rare and often unpublished.

• DeepSeek V4 Pro (CAISI/NIST): The U.S. government's evaluation remains the most notable safety signal, concluding its capabilities were about 8 months behind the frontier .
• GPT-5.5 (AI Security Institute): The UK's AI Security Institute reported GPT-5.5 had a 71.4% pass rate on its expert-level cyber tasks, noting it "may be the strongest model we have tested" on that measure .

How to Think About These Results

It's tempting to crown a single winner, but the data paints a more practical picture:

• For pure coding ability and agentic tasks, Claude Fable 5 is in a class of its own.
• For heavy mathematical reasoning and autonomous command-line work, GPT-5.5 is the tool of choice.
• For a balanced, cost-effective model that's fast and reliable, Gemini 3.5 Flash is a compelling option, though you should validate its coding claims for your specific use case.
• If you need a specialist for legal reasoning, finance, or instruction-following at a low price, Grok 4.3 is a surprisingly potent choice.
• If you're building on a budget and need frontier-grade coding with an open-source license, DeepSeek V4 Pro is unmatched in value, but its real-world sophistication may not match the very best.

Always remember that benchmark conditions vary (prompting, tool use, reasoning effort) and aren't a perfect predictor of your specific application.