AI Models 2026: Release Updates and August Comparison
The benchmark and pricing comparison on this page is an August 14, 2026 snapshot. For a model choice today, start with the release check below, verified against provider documentation on October 5, 2026. The newer releases were not included in the August comparisons; this update does not rerun those tests or declare a new winner.
| Provider | What the current provider documentation lists | Where to read next |
|---|---|---|
| OpenAI | GPT-6 Astra and GPT-6.1 Sol in the API model catalog. | Our GPT-6 production guide covers reasoning, state and long-running tools. |
| Anthropic | Claude Opus 5.5 and Claude Sonnet 5.5, with their own migration requirements. | Read the Opus 5.5 guide and Sonnet 5.5 guide. |
| Stable Gemini 3.8 Flash in the Gemini API documentation. | Our Gemini 3.8 Flash analysis separates coding scores from agent cost. | |
| xAI | Grok 4.7 is documented for the public API. | The Grok 4.7 guide covers state, caching and routing checks. |
| DeepSeek | V4.1 Flash replaces retired Flash routes; V4 Pro continues separately. The current price table lists both. | Check the exact model route and current billing before an elective migration. |
Gemini 4 Argon has a different access boundary: Google describes an initial rollout to selected cyber defenders through Fairwind, with broader access planned. Treat it as restricted access, rather than a universal public replacement for Gemini 3.8 Flash. Our Argon access guide explains that distinction.
These links are a release and migration check, not an independent ranking of the newer models. API documentation also does not establish availability for every consumer plan, region or account. Recheck the exact service and your workload before buying or changing production.
Historical cutoff: the portfolios, scores, prices, examples and takeaways below describe the August 14 comparison. They are retained so the reasoning and evidence remain inspectable; they are not October 5 performance or price claims.
The model race has reached an awkward stage: the rankings can be accurate and still tell you almost nothing about which model to buy.
Claude can lead an aggregate intelligence index. Grok can lead the same evaluator's agent score. GPT can finish coding work with fewer tokens and steps. Gemini can generate more than five times as fast. DeepSeek can cost a fraction as much and let you download the exact weights. None of those facts cancels the others.
The mistake is treating “best AI model” as a single property. In production, the effective product is the model, reasoning level, tool runtime, cache policy, service tier, and data boundary working together. Change one of those and the winner can change with it. A model at max reasoning may gain a few benchmark points while taking ten times longer to show an answer. A cheap token rate may be erased by verbose reasoning, paid tools, retries, or a long-context surcharge. A one-million-token window may be technically available while the agent compacts its working state much earlier.
This comparison uses the newest major models that a normal customer could actually call on August 14, 2026: OpenAI GPT-5.6 Sol, Anthropic Claude Fable 5, xAI Grok 4.6, Google Gemini 3.7 Flash, and DeepSeek V4 Pro 0813. It excludes invitation-only Claude Mythos 5, restricted GPT-5.6 Cyber and Daybreak programs, the announced-but-unreleased Grok 4.7, and Google's forthcoming Gemini 3.5 Pro.
The short answer is that there is no honest universal winner. There is, however, a sensible winner for each kind of work—and a much better way to build a model stack than choosing one logo for everything.
August 14 takeaways — historical scores and prices
“Public” is surprisingly slippery. A company can announce a model without exposing an API, invite a small research group, silently route consumer users through a model, or publish weights without operating a hosted service.
For this article, a flagship qualifies if a developer can obtain it through a self-serve first-party API or generally available product today. Paid access still counts as public; invitation-only access does not. The portfolio sections also include each provider's more practical workhorse because the nominal flagship is often the wrong default.
This is also a dated comparison, not a timeless leaderboard. Artificial Analysis scores, serving latency, provider routes, launch discounts, and benchmark reruns can move. The figures below were checked on August 14, 2026, and every live measurement should be rechecked before a large purchase.
The cleanest way to understand the market is to look at each provider's escalation ladder. A flagship answers “how good can this stack get?” The workhorse answers “what should handle most of my traffic?” The volume tier answers “what can safely fan out?”
| Provider | Frontier lane | Practical default | Volume lane | Useful distinction |
|---|---|---|---|---|
| OpenAI | GPT-5.6 Sol | GPT-5.6 Terra | GPT-5.6 Luna | One Responses tool plane across three price/capability tiers |
| Anthropic | Claude Fable 5 | Claude Opus 5 / Sonnet 5 | Claude Haiku 4.5 | Fable is nominal flagship; Opus handles hardest work; Sonnet is the workhorse |
| xAI | Grok 4.6 | Grok 4.6 High | Grok 4.3 | Native X retrieval and a 200K-token pricing cliff |
| Gemini 3.7 Flash High | Gemini 3.7 Flash Medium | Gemini 3.5 Flash-Lite | Fastest multimodal lane, with a promotional price clock | |
| DeepSeek | V4 Pro 0813 | V4 Flash 0731 | V4 Flash low/non-thinking | Exact MIT weights; queue suitable work off-peak |
GPT-5.6 is a three-tier family with unusually consistent deployment semantics. Sol, Terra, and Luna each expose a 1.05-million-token context window, 128,000-token output, image input, function calling, structured output, and the Responses API's expanding tool surface. That surface includes web and file search, code execution, hosted shell, apply_patch, computer use, MCP, skills, and a multi-agent beta.
Sol is the flagship at $5 per million input tokens and $30 per million output tokens. After OpenAI's July 30 repricing, Terra is $2/$12 and Luna is $0.20/$1.20. Older launch pages and search snippets still show stale Terra and Luna rates, which is exactly why model comparison articles age so badly. Current OpenAI API pricing
The hidden catch is long context. Once input exceeds 272,000 tokens, OpenAI reprices the entire request at twice the normal input rate and 1.5 times the output rate. Sol becomes $10/$45, not merely “a few expensive tokens at the end.” OpenAI's million-token window is real, but it is also a different economic product above that threshold.
Sol's strongest trait is not that it wins every benchmark. It is the combination of frontier agent quality, broad native tools, and relatively disciplined output. On the current independent suite, Sol Max emits fewer than half as many output tokens per task as Claude Opus 5 Max. In the DeepSWE coding harness, Sol and Opus are statistically tied, while Sol uses about half the output and substantially fewer agent steps.
The practical ladder is simple: Luna explores and fans out, Terra handles most reversible production work, and Sol plans, reviews, or takes over when failure is expensive. RohitAI's GPT-5.6 GA analysis goes deeper on the product family and ChatGPT Work layer.
Claude Fable 5 is Anthropic's most capable branded public model. It accepts text and images, supports one million input tokens and 128,000 output tokens, uses always-on adaptive thinking, and costs $10/$50. Yet the best practical Anthropic model for many teams is Claude Opus 5, not Fable. Anthropic model pricing
Opus 5 currently scores higher than Fable on Artificial Analysis, costs half as much at $5/$25, has a more recent May 2026 reliable knowledge cutoff, and can participate in approved zero-data-retention arrangements. Anthropic itself recommends Opus 5 as the starting point for complex agentic coding and enterprise workflows. Sonnet 5, now permanently priced at $2/$10, is the faster workhorse.
Anthropic designates Fable 5 a Covered Model, requiring 30-day retention and making it unavailable under zero data retention. Anthropic's covered-domain safeguards can route some cyber, bio/chem, frontier-model, or distillation conversations to a fallback model. Artificial Analysis labels its result Fable 5 with Opus 4.8 fallback. That score measures a policy-routed service, not a pure checkpoint in isolation.
This does not make Fable bad. It means choosing Fable is also choosing a retention policy, a safety router, and a $50 output rate. Use it when your own evaluation proves an advantage that Opus cannot deliver—not merely because the name sits above Opus in the catalog. RohitAI's Fable 5 and Mythos 5 analysis explains the release boundary in detail.
Grok 4.6 arrived on August 12 and moved xAI squarely into the current frontier cluster. It is publicly callable through the xAI API and powers Grok Build, with distribution through Cursor, Office add-ins, and API gateways. Its model card says consumer web, mobile, and Grok-in-X access will come later.
The model accepts text and images, exposes 500,000 tokens of context, and supports low, medium, high, and xhigh reasoning. Reasoning cannot be turned off. Its strongest product difference is the tool layer: native web search, X search, code execution, collections, remote MCP, function calls, and encrypted reasoning/compaction artifacts.
Below 200,000 input tokens, Grok 4.6 costs $2 input and $6 output per million. At 200,000 or above, the entire request reprices to $4/$12. Priority service doubles those rates again. That makes prompt shape part of cost architecture: a 199K request and a 201K request are not almost the same bill. xAI pricing
Grok 4.6's August 14 independent snapshot is striking. It essentially ties GPT-5.6 Sol on aggregate intelligence and leads this five-model set on Artificial Analysis's Agentic Index and Terminal-Bench 2.1. But “fast frontier model” needs care: median output is around 66 tokens per second, while the first visible answer took roughly 32 seconds in the long-prompt latency test. The model is better suited to background research, planning, and hard agent work than an instant conversational interface unless xAI's priority tier materially changes your traces.
RohitAI's Grok 4.6 analysis covers reasoning levels and the context cliff. xAI subsequently published a detailed Grok 4.6 model card covering coding, search, factuality, cyber, bio/chem, jailbreaks, output safety, child safety, mental health, and model behavior.
Google's most compelling public model today is not the older gemini-3.1-pro-preview. It is the GA Gemini 3.7 Flash, released on August 13.
Gemini accepts text, images, audio, video, and PDFs, supports 1,048,576 input tokens and 65,536 output tokens, and connects to function calling, code execution, File Search, URL context, Search and Maps grounding, caching, and Computer Use Preview. It produces text rather than native image or audio output, and it does not support the Live API.
Its strongest configuration choice is not High; it is Medium by default, High by exception. Artificial Analysis measured the same 45.10 Agentic Index at both levels. Medium cost about 35% less, reached the first chunk 57% sooner, and emitted far fewer tokens. High helps on the hardest knowledge and terminal tests, but routing every call to High wastes Gemini's main advantage.
The launch price is $0.75 input and $3.75 output per million through December 31. Every listed inference and cache rate doubles on January 1, 2027. Google applied the same promotion to 3.6 Flash, so 3.7 is not currently half the adjacent model's live price; it is half the older list price. Gemini API pricing
Gemini's 340-token-per-second median output is in a different speed class from the other flagships. Yet it still took about 9.8 seconds to show the first chunk, and a 100K-token prompt pushed that delay to roughly 16 seconds. Decode throughput and perceived responsiveness are separate product metrics. RohitAI's Gemini 3.7 Flash guide covers the migration and High-versus-Medium economics.
DeepSeek V4 Pro 0813 is generally available through DeepSeek's app, web product, and API under the stable alias deepseek-v4-pro. It offers one million tokens of context, an unusually large 384,000-token output ceiling, low/high/max reasoning, tool calls, JSON output, and compatibility layers for OpenAI Chat Completions, Anthropic Messages, and a partial Responses API adapted for Codex.
It is text-only, and the compatibility story has limits: computer use, MCP, file search, background mode, stored conversations, and several Responses features are unsupported or silently ignored. Wire compatibility is useful; it is not parity. DeepSeek thinking controls · Codex integration limits
The exact DeepSeek-V4-Pro-0813 weights are public under MIT. The checkpoint is about 1.65 trillion parameters and 892.8 GB, using mixed FP4/FP8 storage. DeepSeek's reference vLLM deployment uses one four-GB300 node. This is an enterprise hedge against API and jurisdiction risk, not a laptop-local model—but it is still a form of control none of the other four vendors offers.
Through August 16 at 16:00 UTC, V4 Pro costs $0.003625 cache-hit input, $0.435 cache-miss input, and $0.87 output per million tokens. The new tariff then becomes $0.022/$0.66/$1.98 off-peak or $0.044/$1.32/$3.96 during the 01:00–04:00 and 06:00–10:00 UTC peak windows. Those “half-price off-peak” rates are still substantial increases from the launch tariff. DeepSeek has turned the clock into part of model routing: queued jobs, fan-out agents, and evaluations should know what time it is. DeepSeek API pricing
The RohitAI V4 Pro analysis covers the tariff change, open weights, and the gap between DeepSeek's official Terminal-Bench result and the independent measurement.
Vendor launch tables are useful for learning what a company optimized. They are poor instruments for declaring a universal champion because providers use different prompts, tools, harnesses, sampling rules, and sometimes private test sets. The broadest common snapshot available when this article was checked on August 14, 2026, comes from Artificial Analysis, which runs models through the same index framework and first-party APIs.
Even this is not a laboratory-perfect contest. “Max,” “xhigh,” and “adaptive max” are provider-relative settings, not equal inference budgets. Fable's result includes an Opus 4.8 fallback. Latency reflects specific provider routes at one point in time. The table is best read as a map of trade-offs. Direct snapshots: Fable 5, GPT-5.6 Sol, Grok 4.6, Gemini 3.7 Flash, and DeepSeek V4 Pro.
The table deliberately uses time to first answer, not merely time to first streamed chunk. DeepSeek begins streaming hidden reasoning after roughly 1.85 seconds in this test, but the answer itself arrives around 27.35 seconds.
| Model and tested mode | Intelligence | Agentic | Terminal-Bench 2.1 | Output speed | First answer | AA cost/task |
|---|---|---|---|---|---|---|
| Claude Fable 5 adaptive max + fallback | 62.07 | 56.59 | 84.64% | 62.82 tok/s | 100.42s | $3.140 |
| GPT-5.6 Sol max | 60.93 | 57.78 | 88.01% | 61.73 tok/s | 176.57s | $1.231 |
| Grok 4.6 high | 60.92 | 58.68 | 88.39% | 65.84 tok/s | 32.30s | $0.837 |
| Gemini 3.7 Flash high | 56.03 | 45.10 | 85.77% | 340.07 tok/s | 9.83s | $0.402 |
| DeepSeek V4 Pro max | 53.20 | 49.56 | 78.65% | 78.44 tok/s | 27.35s | $0.252* |
Three conclusions survive the caveats.
First, Fable's final point of aggregate intelligence is expensive. It scores only about 1.1 points above Sol and Grok in this snapshot, yet costs roughly 2.6 times Sol and 3.8 times Grok per Artificial Analysis task. Its route can also fall back to Opus 4.8. That premium may be worth it in a narrow reasoning domain, but it should be proven rather than assumed.
Second, agentic performance and general intelligence have already separated. Grok leads the Agentic Index. DeepSeek beats Gemini on that index even though Gemini is three points ahead on aggregate intelligence. A model that answers hard questions well is not automatically the model that keeps a tool loop on track.
Third, OpenAI's advantage is concision. Sol Max produced about 16,900 output tokens per index task, compared with 35,600 for Fable, 36,800 for Gemini, and 38,900 for DeepSeek. Concision helps cost, reviewability, and agent-loop stability even when token prices are higher.
DeepSWE v1.1 gives a more controlled view of repository work. It runs mini-swe-agent across 113 tasks in 91 repositories and five languages. Confidence intervals overlap heavily, but the cost and behavior differences are illuminating. The scores and costs below are a leaderboard snapshot checked August 14, 2026.
| Model | Resolved | Measured cost/task | Output tokens | Agent steps |
|---|---|---|---|---|
| Claude Opus 5 Max | 74% ±4 | $11.84 | 118K | 99 |
| GPT-5.6 Sol Max | 73% ±3 | $8.39 | 60K | 61 |
| Claude Fable 5 Max | 70% ±4 | $21.63 | 119K | 88 |
| Grok 4.6 xhigh | 67% ±2 | $5.50 | 71K | 87 |
| Gemini 3.7 Flash High | 65% ±2 | $2.18 | 107K | 125 |
| DeepSeek V4 Pro Max | 63% ±6 | $0.06* | 106K | 155 |
Opus and Sol are effectively tied on completion. Sol reaches that result with roughly half Opus's output and 38 fewer steps, which makes it easier to monitor and often cheaper. Fable costs more than either and does not lead this harness. Gemini's giant jump from 3.6 is real, but it takes many more steps than Sol. DeepSeek is dramatically cheaper, yet its 155-step average suggests more opportunities for tool errors, latency accumulation, and brittle recovery.
There is another useful result below the flagship row: GPT-5.6 Luna Max resolves 67% ±4 at about $0.61 per task. That is roughly nineteen times cheaper than Opus for a seven-point difference, with overlapping uncertainty around some mid-pack models. For parallel bug exploration, inventory, test generation, and reversible first passes, a volume model can be the rational choice. Escalate the uncertain cases instead of buying a flagship for every token.
The ugly counterexample is Claude Sonnet 5 Max. Despite its attractive $2/$10 rate, it resolves 54% while costing $26.40 per task in this harness because it emits 214K output tokens and takes 268 steps. That is not a verdict on Sonnet 5. It is a verdict on blindly selecting Max.
Provider labels make reasoning controls look standardized. They are not. OpenAI Max, Anthropic Max, Grok xhigh, Gemini High, and DeepSeek Max do not represent the same compute budget or latency target. They are policy knobs inside different systems.
OpenAI offers none through max, with medium as the default. Claude Fable 5, Opus 5, and Sonnet 5 use adaptive thinking; Fable's adaptive thinking is always on, while Haiku 4.5 does not support it. Grok 4.6 cannot disable reasoning. Gemini supports low, medium, and high but removed minimal. DeepSeek supports low, high, and max; requests for medium or xhigh map to high.
The practical cost of that knob is visible in like-for-like provider tests:
This creates a better operating rule: Medium or High is the normal production SKU; Max is an exception budget. Route Max only when the task is ambiguous, irreversible, or expensive to fail—and only after owned traces show that the added reasoning changes the acceptance rate.
Simple $/million-token tables hide the most consequential part of 2026 pricing: each provider has attached cost to a different dimension.
OpenAI charges more after a context threshold. Anthropic charges for a premium model and cache lifetime, but keeps its one-million-token rate flat. xAI also has a context threshold, with priority service doubling the bill. Google has a calendar cliff when its launch promotion expires. DeepSeek has time-of-day pricing. The same workload can change price without changing a single prompt word.
| Flagship | Normal input / output | The catch | Useful response |
|---|---|---|---|
| GPT-5.6 Sol | $5 / $30 | Above 272K input, entire request becomes $10 / $45 | Retrieve, summarize, or split state before crossing the threshold |
| Claude Fable 5 | $10 / $50 | Cache writes cost 1.25× or 2×; reads cost 0.1× | Use Fable only where task uplift beats Opus and cache economics |
| Grok 4.6 | $2 / $6 | At 200K input or above, entire request becomes $4 / $12 | Track prompt size before tool output pushes the run over |
| Gemini 3.7 Flash | $0.75 / $3.75 promo | Every listed rate doubles January 1, 2027 | Model steady-state economics now, not after integration |
| DeepSeek V4 Pro | $0.66 / $1.98 off-peak* | Peak is $1.32 / $3.96 during two daily UTC windows | Queue evaluation, research, and CI fan-out off-peak |
For a deliberately simple 1,000-token input plus 500-token output request—with no cache, tools, hidden reasoning, retries, long-context surcharge, or service-tier premium—the bare token floor is:
That arithmetic is useful only as a floor. Gemini Search grounding can cost $0.014 per call after its 5,000-request shared allowance—more than five times the example's model tokens. xAI's hosted web, X, and code tools cost $0.005 per invocation, and xAI lists no free allowance. High reasoning emits more billable tokens. A cheaper model that takes twice as many steps can lose the task-level comparison.
The right financial metric is:
cost per accepted result =
model tokens
+ cache writes and storage
+ paid tool calls
+ failed attempts and retries
+ human review
+ cost of escaped errors
Per-token price is procurement metadata. Cost per accepted result is product economics.
Four of the five flagships advertise roughly one million input tokens; Grok 4.6 offers 500,000. That makes the raw context number look settled. It is not.
Context length answers one narrow question: how much can the endpoint admit? It does not tell you how well the model retrieves a detail in the middle, how long prompt prefill takes, whether the agent compacts state early, how cache storage is billed, or whether tool results remain semantically useful after dozens of turns.
The differences are substantial:
Production agents should track effective working state, not advertised context. Log raw prompt tokens, retrieved tokens, cache hits, compaction events, discarded evidence, and the exact summary handed to the next turn. A compacted 130K working set can outperform a sloppy one-million-token transcript.
Scores are clustering, so providers are competing with their execution systems.
OpenAI's moat is the broadest unified agent surface. The Responses API can combine hosted search, files, shell, code, patches, computer use, MCP, skills, and multi-agent work under one protocol. Anthropic's moat is Claude's coding behavior, Messages API, MCP ecosystem, and a practical Opus/Sonnet ladder—although specific tool and priority support varies by model.
xAI's moat is live retrieval, especially X search, plus Grok Build and an API that exposes encrypted reasoning and compaction. Google's moat is the complete multimodal lane: video, audio, PDFs, Search, Maps, code execution, caching, Antigravity, Workspace, and extremely fast decoding. DeepSeek's moat is portability: compatible API surfaces, its own harness, and downloadable weights that provide a genuine exit option.
This is why switching providers is never just changing model=. The real execution lockfile is:
resolved model version
+ reasoning policy
+ system prompt and tool schema
+ agent harness and retry rules
+ cache and compaction behavior
+ service tier and regional route
+ retention and data-use policy
Google's Antigravity provides the clearest warning. The outer ID antigravity-preview-05-2026 moved from Gemini 3.6 to 3.7 after only 23 days. OpenAI's gpt-5.6 is a mutable alias that routes to Sol, while Anthropic's newer dateless Claude IDs are documented as pinned snapshots. A team that records only the friendly product name may be unable to reproduce yesterday's run.
For high-stakes agents, attach an execution manifest to every trace: provider, exact resolved model, effort, tools, tool versions, cache key, context compaction, client SDK, timestamp, region, service tier, and policy route when observable. This sounds bureaucratic until a silent model change breaks an evaluation or an audit asks which system made a decision.
Capability is irrelevant if the data boundary disqualifies the service. The five providers do not offer interchangeable privacy contracts.
| Provider | Commercial API training | Retention/control caveat | What to verify |
|---|---|---|---|
| OpenAI | Not used for training absent opt-in | Default abuse logs up to 30 days; Responses stores state by default; approved ZDR has feature limits | Endpoint eligibility, state storage, files, batches, region |
| Anthropic | Not used absent opt-in or feedback | Opus/Sonnet may be ZDR-eligible; Fable requires 30-day retention and is not | Covered-model routing, batch/files/MCP retention |
| xAI | API data not used without explicit permission | Normally retained up to 30 days; where available, ZDR disables stateful Responses, Files/Collections, and Batch | Consumer versus API policy, ZDR eligibility, and lost features |
| Paid API data not used for product improvement | Outside EEA/Switzerland/UK, free AI Studio/API data may be reviewed; interaction state and grounding have separate rules | store setting, Search/Maps retention, region, paid project route | |
| DeepSeek | No explicit public ZDR promise found | Consumer data can improve models and is processed/stored in China; exact open weights enable self-hosting | Jurisdiction, downstream consent, hosted versus owned deployment |
Open weights do not automatically mean stronger safety transparency. DeepSeek published the exact 0813 checkpoint, yet its public safety material remains tied to the broader April V4 family; no new GA-specific red-team report was found. xAI published a 36-page Grok 4.6 model card on August 12. Google published a 3.7 model card but said the complete Frontier Safety Framework report would follow. OpenAI published a 5.6 system card. Anthropic published detailed Fable safeguards, while a dedicated Opus 5 card was not visible in its public index when checked.
The source documents are worth reading before procurement: OpenAI data controls, Anthropic API retention, xAI API security and retention, Gemini API terms, DeepSeek's privacy policy, and its Open Platform terms. Consumer plans and commercial APIs can follow different rules.
The point is not to award a moral score. It is to separate four ideas that marketing often blends: weight access, training-data disclosure, safety evaluation, and contractual data control. A provider can be strong on one and weak on another.
Opus 5 High leads Sol High on the current Artificial Analysis aggregate, while Opus 5 Max and Sol Max are statistically tied on DeepSWE. Sol is more token- and step-disciplined in that coding harness and offers the broader unified tool plane.
The best fit when current web evidence and native X retrieval matter. Budget for slow first answers, watch the 200K context cliff, and verify important claims outside one retrieval channel.
The strongest default for video, audio, PDFs, Search-grounded tasks, and high-throughput agents. Use High as an escalation lane and model the January price doubling now.
Use Flash for routine fan-out and Pro for difficult verification. Schedule externally queued work off-peak after August 16. Choose self-hosting only when jurisdiction or platform control justifies serious infrastructure.
Luna's independent coding economics are unusually strong. Let cheap workers inspect, test, classify, and propose; send uncertain or irreversible decisions to Sol or Opus.
Start with retention, residency, ZDR, storage, and connector requirements. Fable's mandatory retention and DeepSeek's hosted jurisdiction can remove them before capability testing begins.
If one default has to cover a broad builder workload, my practical shortlist is GPT-5.6 Terra, Claude Opus 5, and Gemini 3.7 Flash Medium, not the five most expensive endpoints. Terra has the unified OpenAI agent plane, Opus has excellent hard-task behavior without Fable's policy baggage, and Gemini has speed plus the broadest multimodal input surface. Grok becomes the default when X-native research matters. DeepSeek becomes the default when price or control dominates.
Do not evaluate with ten trivia prompts and a vibes score. Build a small trace set from work your team actually performs, then measure the cost of acceptance.
Keep three scores per configuration:
A model can win the first and lose the other two. That is often the right trade for architecture decisions or a final security review. It is usually the wrong trade for 500 parallel repository scans.
My read is that model competition will become less legible before it becomes simpler.
First, flagship names will matter less than runtime manifests. Managed agents will silently change models, providers will tune routes, and reasoning policies will evolve faster than version numbers. Serious teams will version the whole execution contract.
Second, the middle tier will eat more flagship traffic. Terra, Opus, Gemini Medium, Grok High, Luna, and DeepSeek Flash are getting good enough that the expensive model will become an escalation specialist. The winning provider may be the one with the best router, not the highest single score.
Third, price will become programmable. Time-of-day queues, threshold-aware context trimming, cache-density routing, batch or externally queued lanes, and tool budgets will sit beside prompts in agent infrastructure. DeepSeek's clock, Google's calendar, and xAI/OpenAI context cliffs are early versions of the same idea.
Fourth, open weights will become an insurance product. Most companies will not self-host an 893 GB model cheaply. Some will still value the credible ability to leave a hosted API, keep a regulated workload inside a boundary, or audit a pinned artifact. The option has strategic value even when the servers remain off.
Finally, independent agent evaluation will get harder. Providers increasingly optimize model and harness together. DeepSeek reports 87.9 on Terminal-Bench while Artificial Analysis measures 78.65; that nine-point gap is a reminder that harness, effort, tools, and fallback belong beside every score. “The model got 88” is no longer a complete sentence.
If I had to summarize the public frontier in one line each:
The best AI model in 2026 is not one endpoint. It is a routing policy that knows which endpoint deserves the next task.
There is no universal winner. Claude Opus 5 currently leads the broad independent aggregate among practical public models, GPT-5.6 Sol is exceptionally strong for coding and integrated agents, Grok 4.6 leads the current Agentic Index snapshot, Gemini 3.7 Flash wins throughput and multimodal breadth, and DeepSeek V4 Pro wins cost and downloadable control.
Fable scores slightly higher on aggregate intelligence in the current Artificial Analysis snapshot, but Sol leads it on Agentic Index, Terminal-Bench, token discipline, price, and the breadth of its unified agent tools. Fable's result also includes an Opus 4.8 fallback. The better choice depends on the task and retention requirements.
Start with Opus 5. It currently scores higher independently, costs half as much, is Anthropic's recommended complex-work model, and can support approved zero-data-retention arrangements. Test Fable only where its own-task uplift justifies mandatory 30-day retention and possible safety fallback.
Claude Opus 5 and GPT-5.6 Sol are statistically tied at the top of the current DeepSWE v1.1 board. Sol uses fewer output tokens and agent steps. Grok 4.6, Gemini 3.7 Flash, and DeepSeek V4 Pro form a lower-cost cluster whose confidence intervals overlap; the right choice depends on tools, latency, and budget.
Gemini 3.7 Flash has the fastest measured decoding in this group at about 340 output tokens per second. That does not mean instant responses: the same test measured roughly 9.8 seconds to the first chunk, and long inputs can increase prompt-processing delay.
DeepSeek V4 Pro is the cheapest named flagship on raw token pricing off-peak and remains lowest in this article's recalculated Artificial Analysis task arithmetic. During DeepSeek's peak windows, promotional Gemini 3.7 Flash has slightly lower raw input and output rates. For large fan-out, DeepSeek Flash, GPT-5.6 Luna, and Gemini Flash-Lite deserve separate evaluation.
GPT-5.6, Claude Fable/Opus/Sonnet 5, Gemini 3.7 Flash, and DeepSeek V4 support roughly one million input tokens. Grok 4.6 supports 500,000. The advertised limit does not guarantee equal recall, latency, cache economics, or uninterrupted agent memory.
Gemini 3.7 Flash has the broadest input mix here: text, images, audio, video, and PDFs, plus Google Search and Maps grounding. GPT-5.6, Claude, and Grok accept images in addition to text. DeepSeek V4 Pro is text-only.
DeepSeek V4 Pro 0813 is the only flagship in this comparison with exact downloadable MIT-licensed weights. “Open weight” is more precise than “open source” for a model checkpoint. Its roughly 893 GB footprint makes production self-hosting an infrastructure project.
The benchmark comparison includes only models publicly callable on August 14, 2026. At that cutoff, Mythos was limited-access, while Grok 4.7 and Gemini 3.5 Pro had not been publicly released. Grok 4.7 now appears in the dated release check above; it has not been retroactively added to the August benchmark table.
Begin with data and tool requirements, then evaluate 30–50 real tasks at sensible reasoning levels. Measure accepted completion, latency, agent steps, token/tool spend, human review, and failure severity. Use the results to create an escalation router instead of forcing every workload through one provider.