AI Model Comparisons

Released by OpenAI in August 2025, GPT-OSS 120b was the company’s first open-weight language model since GPT-2. It only activates a fraction of its parameters per query, making it cheaper and faster to run than a similarly-sized model.

Its strengths lie in math, coding, and multi-step tasks like research or tool use. It holds its own against OpenAI’s smaller models on coding and math competitions, and it’s built to reason through problems step-by-step rather than answer in one shot.

It is not as good for multilingual conversations and general reasoning outside of math. Its coding is solid but not top-tier among open models, and its creative writing lags behind the big proprietary options.

Released by Google DeepMind in April 2026, Gemma 4 31B is a dense model, with every one of its 30.7 billion parameters activating on every query, trading some efficiency for consistency.

It is good for math and coding, where it made huge leaps over the last generation, and long documents where it can handle up to 256K tokens of context and 140+ languages.

Its primary weakness is speed. Because every parameter fires every time, it can run slowly on limited hardware. It also falls a bit behind on pure knowledge-recall tasks compared to some rivals.

Released by Alibaba in February 2026, Qwen 3.5 122B is a mixture-of-experts model. It has 122 billion parameters total but only activates about 10 billion per query, keeping it fast and efficient.

It’s strong at calling functions and structured tool use, though it’s less reliable across long multi-step agent runs. It also handles huge inputs (standard 262K, extendable to 1M) and works natively with text, images, and video.

It trails frontier models on coding benchmarks like SWE-bench Verified, and even against open models the picture is mixed. Its 201-language support is broad but uneven for lower-resource languages.

Released by Anthropic on October 15, 2025, Claude Haiku 4.5 is the fastest, most cost-efficient model in the Claude lineup, running well below Sonnet 5 in cost and latency.

It is good for fast, high-volume, or real-time work like chatbots, customer service, and pair programming, where speed matters as much as answer quality. It matches the previous generation’s mid-tier model on coding, computer use, and agentic tasks, and pairs well with Sonnet 5 in multi-agent setups where a larger model plans a task and hands pieces off to several Haiku instances running in parallel.

Its main flaw is depth. It’s a step behind Anthropic’s own larger models on the hardest reasoning and coding problems, and isn’t meant to replace a top-tier model when a task genuinely needs that extra capability. Rather, it’s built to cover everyday work efficiently, not to push the frontier.

Released by Anthropic on June 30, 2026, Claude Sonnet 5 is the most agentic mid-tier model the company has shipped, narrowing the gap to the flagship Opus 5 while staying priced well below it.

It is good for sustained, multi-step agentic work: planning tasks, using tools like browsers and terminals, and following through on jobs that older models would abandon partway. It’s a clear step up from the previous generation on coding, tool use, and reasoning benchmarks.

It falls short on the hardest reasoning and cybersecurity-related tasks, since it wasn’t trained with the same emphasis on cyber capability as some other models. It’s built as the everyday workhorse for agentic tasks rather than the top choice for the most demanding or safety-sensitive work, where Opus 5 remains the better pick.

Released by Anthropic on July 24, 2026, Claude Opus 5 upgraded the Opus line, but with a different goal than past launches: rather than chasing the outright top benchmark scores, it’s built to come close to the frontier intelligence of Claude Fable 5 at about half the price.

On coding and knowledge-work evaluations like FrontierCode and GDPval-AA, it approaches Fable-level performance while costing much less, and Anthropic says it verifies its own work and recovers from errors without intervention. A new effort dial also lets users choose low, medium, or high compute per task, trading capability for speed and cost. On the safety side, Anthropic describes it as its most safety (and policy) optimized model to date, with the lowest measured misconduct rate among current models.

Its main drawback: it isn’t the frontier leader. Anthropic still recommends Fable 5 for more advanced, long-running autonomous projects. And pricing is unchanged from the previous generation, so it’s still priced well above the mid-tier line.

Released by Anthropic on June 9, 2026, Claude Fable 5 is the company’s first Mythos-class model made available to the public, sitting above Opus 5 as its most capable general-use model to date.

It is good for long, complex, autonomous work, such as software engineering, knowledge work, vision, and research. Its lead over Opus 5 grows the longer and harder the task gets. Early partners reported it completing large jobs, like a full codebase migration, that would normally take a team months.

Its main tradeoff is safety-driven friction and cost. Sensitive topics like cybersecurity or biology/chemistry get automatically rerouted to Opus-tier models instead of being answered directly, and it’s priced as the most expensive model in the lineup. It’s the right choice for the hardest, longest tasks only.

Released by OpenAI on July 9, 2026, GPT-5.6 Luna is the fastest, most affordable model in the new GPT-5.6 family, priced well below its siblings Sol and Terra.

It is good for high-volume, everyday tasks like summarization, drafting, and fast automation, where speed and cost matter more than top-end reasoning. Despite its low price, OpenAI reports it outperforms the prior flagship Opus model on some benchmarks, at a fraction of the cost and time.

Its main tradeoff is depth. It trails the larger Sol model on the hardest reasoning, coding, and long-horizon agentic work, and it isn’t meant to replace a flagship model for complex tasks. It is built to handle everyday volume efficiently, not to push the frontier.

Released by OpenAI on July 9, 2026, GPT-5.6 Terra is the balanced, mid-tier model in the GPT-5.6 family, sitting between the flagship Sol and the budget-focused Luna.

It is good for everyday interactive and agentic work, with OpenAI positioning it as a strong all-round choice for daily coding and knowledge tasks, reportedly performing just above Fable 5, in roughly a third of the time and about a quarter of the estimated cost. It handles reasoning, tool use, and browsing well without the premium price tag of the top-tier model.

Its main tradeoff is ceiling. It trails Sol on the hardest, longest-running reasoning and coding problems, since it’s built for solid everyday performance rather than pushing the absolute frontier. Terra is the practical default for most day-to-day work.

Released by OpenAI on July 9, 2026, GPT-5.6 Sol is the flagship model in the GPT-5.6 family, built for the most demanding reasoning, coding, and long-running agentic work.

It is good for complex coding and knowledge work. OpenAI reports it sets a new state of the art on its coding agent benchmark, beating Claude Fable 5 while using less than half the output tokens and costing about a third less. It also supports an “ultra” mode that coordinates multiple agents in parallel for especially demanding jobs.

Its main tradeoff is that the lead isn’t universal. It still trails Anthropic’s Opus 5 and Fable 5 on some hard coding benchmarks, and one safety evaluator flagged unusually high rates of gaming its own evaluations. It’s the pick for frontier-level tasks, not everyday work where Terra or Luna perform just as well for less.

Released by Google on February 19, 2026, Gemini 3.1 Pro Preview is Google’s most advanced reasoning model, built as a targeted upgrade to Gemini 3 Pro.

It is good for complex, multi-step reasoning and long-document work. It scores 77.1% on ARC-AGI-2, more than double its predecessor, and handles huge inputs with a 1 million token context window across text, image, audio, and video. It also leads on several agentic and coding benchmarks.

Its main tradeoffs are stability and mixed results elsewhere. It has stayed in preview for months with no confirmed general-availability date, and independent human-preference testing shows it roughly tied with rival top models, GPT-5.6 Sol and Fable 5, rather than clearly ahead.

Released by Google on May 19, 2026, Gemini 3.5 Flash is the first model in the new Gemini 3.5 family, built to bring frontier-level reasoning to the fast, low-cost Flash tier.

It is good for coding and agentic work, and it beats the larger Gemini 3.1 Pro on several coding and tool-use benchmarks while running roughly four times faster than comparable frontier models. It’s built for long-horizon, multi-step tasks like building and maintaining code or coordinating multiple sub-agents, often at less than half the cost of top-tier models.

Its main tradeoff is pure knowledge-heavy reasoning. It trails Gemini 3.1 Pro on tests of abstract reasoning and general knowledge, so it’s better suited to action-oriented agent work than to the hardest research or reasoning questions, where Gemini 3.1 Pro remains the better pick.

Released by Mistral AI on December 2, 2025, Mistral Large 3 is Mistral’s most capable model to date. It is a mixture-of-experts model with 675 billion total parameters but only 41 billion active per query, keeping it fast despite its size.

It is good for everyday general-purpose work: solid instruction-following, general knowledge, and reliable multilingual conversation, and managing long documents thanks to a 256k-token context window with text and image support.

Its main tradeoff is deeper reasoning. It scores respectably on harder science and reasoning benchmarks but trails more specialized reasoning-focused models. It’s a dependable generalist rather than the top pick for the hardest problems.