New artificial intelligence models are appearing almost monthly, claiming to be the most advanced. For users who simply want to keep up with technological developments or need a tool for drafting emails, correcting code, or generating images from text prompts, the information noise becomes excessive.
This is the first edition of the AI Barometer from TechCentral—a regular, fact-based report that shows which models and platforms are leading in the tasks users apply in practice. The report does not rely on vendors' marketing claims but is based exclusively on public, independently conducted benchmarks and leaderboards specifically chosen to test relevant use cases.
The evaluations were checked in the first week of July 2026, and given the pace of new product releases, some results may change by the time of reading. This is why this barometer exists.
Two types of evidence feed into the report. The first is arena-style testing: platforms like Arena (formerly known as LMArena, a research division of the University of California, Berkeley) and Artificial Analysis show users two anonymized results for the same prompt and ask them to choose the best. Millions of anonymous votes are converted into Elo ratings, similar to chess.
The second method is task benchmarking, where models are evaluated against fixed sets of complex, realistic problems, such as debugging real software errors, answering expert-level scientific questions, or creating documents and spreadsheets for which professionals are paid. Both approaches have their drawbacks: votes in Arena reward what people like, not necessarily what is correct. Task benchmarks can be compromised if test questions appear in the model's training data, so the barometer favors contamination-resistant tests. When leading models are within a few points of each other, the honest conclusion is a statistical tie, and this is indicated.
The broadest indicator is presented in the Arena text ranking, which collected 7.15 million anonymous votes among 369 models in the first week of July. The newest model, Claude Fable 5 from Anthropic, leads the ranking with an Elo score of 1,509, with the surprising fact that all five top spots are held by Anthropic models.
The best of the rest is Google's Gemini 3.1 Pro Preview with a score of 1,486, closely followed by OpenAI's GPT-5.5 at 1,481. However, it should be noted that the top 10 are within about 30 Elo points, so any of them will suffice for daily use. Price is more important than quality here: according to Arena's published API tariffs, Gemini 3.1 Pro costs $2 per million input/output tokens (approximately 33/195 rupees), while Claude Fable 5 costs $10/50 (approximately 163/814 rupees), making Gemini an obvious choice for price/quality ratio. This comparison applies to businesses using model APIs; for end consumers, flagship subscriptions from OpenAI, Google, and Anthropic cost within a few dollars of each other, regardless of the base token cost.
For AI performing real work—creating documents, presentations, spreadsheets, and analytical reports in office life—the most relevant public test is GDPval. This set of 220 tasks was created by OpenAI in collaboration with industry specialists from 44 professions. Artificial Analysis runs models through it in agent mode, giving them access to the web and command line, and ranks the results via anonymous pairwise comparison. Anthropic's Claude Opus 4.8 confidently leads with an Elo score of 1,890, surpassing GPT-5.5 by 121 points with a score of 1,769—a remarkable result considering the benchmark was created by OpenAI.
The coding barometer indicator is SWE-bench Pro, which functions as an independent public ranking from Scale AI. It requires models to solve real problems from actively maintained software repositories and was specifically designed to counteract training data contamination, a known issue with the older SWE-bench Verified test. In the independent ranking, OpenAI's GPT-5.4 leads with a score of 59.1%, tied statistically with Meta's recently added Muse Spark at 55%, while Anthropic's Claude Opus 4.6 takes third place with 51.9%. It is worth noting Anthropic's claim that their new model Claude Fable 5 scored 80.3% on the same set of tasks, but this figure was obtained using the company's proprietary tools and has not yet appeared in the independent ranking. Until then, the confirmed leader belongs to OpenAI.
Assessing writing quality is traditionally difficult, and the most rigorous public approach is the independent benchmark EQ-Bench Creative Writing (version 3). It combines rubric scoring with Elo-style pairwise judging across 32 different prompts. Claude Fable 5 leads with an Elo score of 2,192, slightly ahead of Claude Opus 4.7 with 2,179, and GPT-5.5 takes third place with 2,019. There is an important caveat here: the judge is the AI model itself, so the results are biased, not definitive, although the benchmark also publishes raw results allowing skeptical readers to evaluate the prose themselves.
Research is divided into two competencies. For deep reasoning, the standard test is Humanity’s Last Exam, consisting of 2,500 questions compiled by subject matter experts precisely because advanced models could not answer them. In the official Scale AI-managed ranking, Google's Gemini 3.1 Pro Preview leads with a score of 46.4%, tied statistically with OpenAI's GPT-5.4 Pro at 44.3%—meaning even the best models cannot handle more than half of the exam. (Anthropic reports a higher score of 53.3% for Claude Fable 5, but this assessment is provided by the vendor and has not yet been included in the official ranking).
For real-time web research, the benchmark is BrowseComp—an OpenAI test consisting of 1,266 questions that checks the agent's ability to find hard-to-reach information online. Note on objectivity: unlike all other tests in this report, BrowseComp does not have an independently conducted public ranking—labs report their own assessments. According to this self-reported data, OpenAI's GPT-5.5 Pro leads with 90.1%. Thus, for depth—Google, and for finding information on the internet—ChatGPT, according to its own data.
Image generation models are ranked almost entirely based on human preference, and the most active public ranking is Image Arena from Artificial Analysis. OpenAI's GPT Image 2 (high) leads with an Elo score of 1,339 based on over 13,000 comparisons, surpassing second place Reve 2.0 (1,281) by 58 points, while Microsoft's MAI-Image-2.5 takes third place with 1,271.
Video is a category where common perceptions become outdated fastest. In the text-to-video Arena by Artificial Analysis (with audio), ByteDance's Dreamina Seedance 2.0 leads with a score of 1,224, followed by Alibaba's Wan 2.7 (1,160) and HappyHorse 1.1 (1,154). The strongest Western participant, Google's Veo 3.1, is only in tenth place with 1,095. In the silent (no sound) ranking, Alibaba's HappyHorse 1.0 leads with a score of 1,290. Developers from China occupy four spots in the top, and the cost remains high: producing a minute of footage in 1080p resolution costs approximately $9 to $24 (about 145 to 390 rupees) via the creators' own APIs.
Music benchmarking is the youngest and least extensive category, so these results should be considered indicative. In Artificial Analysis's Music Arena, which uses the same anonymous voting method as its graphics and video arenas, Suno v5.5 leads both rankings—instrumental (Elo 1,194) and vocal (1,168)—surpassing Mureka V8 in both cases, while the earlier version Suno V5 takes third place in each table.
