xAI Grok 4.3 and Grok 4.20 in 2026: Context, Cost, and Real Use Cases

What happened

xAI’s Grok family in May 2026 covers two production model tiers with distinct positioning: Grok 4.3 as the current frontier reasoning model, and Grok 4.20 as the long-context multi-agent variant for harder problems ¹ . The xAI May 15 2026 model-retirement notice records that the previous Grok Code Fast 1 tier and seven other legacy slugs were deprecated; requests to those slugs now redirect to grok-4.3 and bill at grok-4.3 rates ² . The API at api.x.ai is the developer entry point, and the consumer Grok inside X (the platform formerly known as Twitter) is the same family wearing a different surface ³ . For a developer evaluating where Grok fits among ChatGPT, Claude, and Gemini, the answer is narrower than the marketing suggests but genuinely useful at the edges. Grok is a strong third or fourth option, not a default; pick it for X-platform data workloads, very-long-context workloads where Grok 4.20’s 2-million-token window helps, or vendor-diversification, and pick Claude Sonnet 4.6, GPT-5.5, or Gemini 2.5 Pro for everything else.

What Grok actually is, in 2026

Grok is xAI’s family of frontier large-language models, started in November 2023 with Grok 1, with point releases through Grok 4.3 by mid-2026 ⁴ . xAI is Elon Musk’s AI lab, and the company’s pitch has been openness about training and a “maximally curious” model posture. The technical reality matters more than the framing: Grok 4.3 is a competitive frontier model on standard benchmarks that xAI now positions as the migration target for developers who previously routed agentic-coding traffic to Grok Code Fast 1 ² .

The lineup as of May 2026 has two buckets a developer cares about. Grok 4.3 is the general-purpose reasoning model with multimodal video, three reasoning-intensity levels, and a 1-million-token API context window, and xAI’s own migration guidance points existing Code-Fast-1 workloads at grok-4.3 for agentic coding and web-development work ² . Grok 4.20 is the long-context multi-agent variant with a 2-million-token context window and a multi-agent runtime that scales from 4 agents at low / medium reasoning effort up to 16 at high / extra-high effort, used when the workload is hard enough that letting the model debate itself improves answers ⁵ . Both are reachable via the api.x.ai endpoint with OpenAI-compatible request shapes, which is the practical reason a team can swap them in without rewriting client code. Legacy slugs including grok-code-fast-1, grok-3, and grok-4-0709 still resolve, but they auto-route to grok-4.3 and bill at grok-4.3’s rate card ² .

Image: xAI Models documentation, used for editorial coverage of the Grok model catalogue and pricing surface.

Pricing for a dev team

Grok’s API uses standard per-million-token billing in USD, with input and output priced separately. As of 2026-05-19, the live rate card on docs.x.ai is the only source to trust at any given moment, and prices fluctuate; verify before purchase ¹ . The structural shape of the pricing is what matters for budgeting more than any specific dollar figure quoted from a single point in time.

The most important practical wrinkle for developers in India is billing access. As of May 2026, xAI’s billing surface does not directly accept Indian payment cards for API service ⁶ . Developers in India route around this through international cards issued by Indian banks where the issuer permits foreign-currency online payments, third-party API resellers that aggregate xAI access onto INR-friendly billing, or by waiting for direct India payment support to land. This is a genuine friction point and worth discovering before committing engineering time to a Grok-routed workload.

The structural shape of the pricing is what matters for budgeting. Grok 4.3 sits in the frontier tier next to GPT-5.5 and Claude Sonnet 4.6, in the premium reasoning band where you expect a few dollars per million input tokens and noticeably more per million output tokens. Grok 4.20 charges more because the multi-agent runtime fans out compute across 4–16 agents (scaling with reasoning effort) and its 2-million-token context window. The cheap-and-fast tier xAI previously occupied with Grok Code Fast 1 is no longer a separate billing surface: legacy code-fast traffic now bills at Grok 4.3 rates after the May 15 2026 retirement ² . The economics of a solo developer in India therefore look different from the pre-retirement picture: cost-per-token for coding-agent loops has effectively risen to the Grok 4.3 frontier-tier rate, which makes the cheap-and-fast comparison now point at GPT-5.5 Mini, Claude Haiku 4, or Gemini 2.5 Flash for budget-bound routing rather than at any xAI-only tier.

GST treatment of foreign API spend follows the standard pattern that already applies to OpenAI, Anthropic, and Google API invoices billed to a registered Indian business; consult your accountant for the exact rate that applies to your billing arrangement and your business’s tax registration status. Treat the GST surface as a known cost line, not a gotcha.

Where Grok wins

Two lanes are genuinely Grok’s territory in 2026, plus a third diversification lane. Being honest about which lane your workload sits in saves a lot of switching cost.

X-platform data integration. Grok inside X has access to real-time X posts as a search surface via xAI’s Live Search API, a feature no other frontier-model vendor offers natively ⁷ . For workloads that involve sentiment analysis on Twitter chatter, real-time event tracking (cricket scores, election results, product launches), or aggregating opinions from X discussions, Grok’s X integration is the only path that doesn’t involve scraping, the X firehose API at enterprise pricing, or third-party social-listening platforms. This is the place where Grok’s value is hardest to reproduce.

Very-long-context multi-agent reasoning. Grok 4.20’s 2-million-token context window paired with its multi-agent debate runtime is a genuinely differentiated shape for workloads that need to load a large corpus into context and have multiple reasoning passes contest the answer. Use cases where this lane bites: large-codebase audits where you’d rather load the whole repository than chunk it, multi-document legal or research synthesis above a million tokens, and adversarial evaluation flows where you want multiple agents to argue against each other before producing a final answer. Gemini 2.5 Pro matches the raw context size but does not ship a comparable multi-agent runtime on the same API surface.

Vendor diversification. Teams built on OpenAI or Anthropic have one more thing to think about: what happens when that vendor changes its terms, raises prices, or suspends accounts. Grok is the credible third frontier vendor alongside Google’s Gemini, and its OpenAI-compatible API surface makes it cheap to wire in as a fallback path in your model-routing layer. You don’t have to use Grok as your primary; the work to add it as a backup is hours, not weeks.

Image: xAI API reference, used for editorial coverage of the OpenAI-compatible API surface and tool-calling features.

Where Grok loses

The honest cases where Grok is not the right pick are larger than the wins.

Speed-bound coding workflows. The previous version of this article called speed-bound coding a Grok lane on the strength of the Grok Code Fast 1 tier. With Code Fast 1 retired and traffic routed to Grok 4.3 ² , the cheap-and-fast economics that made the lane attractive no longer hold on xAI’s surface. The aggregated source consensus now points elsewhere for that workload: GitHub Copilot, Cursor, Cline, and Claude Code all route to combinations of Claude Sonnet 4.6, Claude Haiku 4, GPT-5.5 Mini, or Gemini 2.5 Flash for high-volume agentic-coding loops where tokens-per-second and per-call price matter as much as quality ^{8 9} . Teams previously routing agentic-coding traffic to Grok Code Fast 1 should evaluate one of those alternatives before defaulting to grok-4.3 at the frontier rate.

For coding-quality-first work (multi-file refactors, full-stack TypeScript or Python projects, agent loops where tool-call reliability matters more than raw throughput), Claude Sonnet 4.6 leads independent benchmarks and developer surveys ⁹ . For agent-mode work where tool reliability matters (long-running tasks, multi-step workflows, structured-output discipline), GPT-5.5 and Claude have a longer track record and broader tool-ecosystem support ¹⁰ . Grok’s tool-calling has improved across its versions but it is not the ecosystem leader.

For long-document RAG and long-context work above a few hundred thousand tokens, Gemini 2.5 Pro’s pricing structure remains the cheapest path among the frontiers. Grok 4.20’s 2-million-token context window matches Gemini on raw size, but its per-token price for very long inputs does not undercut Gemini in the way Gemini undercuts everyone else.

For multilingual work in Hindi, Tamil, Telugu, or Bengali (coding-comment generation, doc-translation workflows), the picture is mixed. All four frontier vendors handle Hindi reasonably; none of them is dramatically better. If multilingual quality is the binding constraint, run a bake-off rather than trusting any vendor’s marketing claim.

What to test, in 30 minutes

If the question is “should we add Grok to our stack,” the cheapest experiment is a 30-minute bake-off. Sign up at x.ai/api, sort out the billing path that works for your card situation, get an API key, and run three concrete tests against the prompts you’d send to your current vendor.

First, the X-data feature: if your product touches social listening at all, test Grok’s Live Search integration on a question your current stack can’t easily answer. “What are people in Mumbai saying about a specific consumer brand this week” is a representative shape; see if the answer is useful and timely.

Second, long-context multi-agent reasoning on Grok 4.20: pick a workload large enough to need north of one million tokens of context (a full mid-sized repository, a multi-document legal bundle, a research-paper synthesis with all citations), run it at the multi-agent setting your reasoning-effort budget allows, and compare the final answer’s quality against your current long-context vendor. The 2-million-token context plus 4–16-agent debate is the differentiator; if your workload does not need it, the lane does not bite.

Third, agent reliability on Grok 4.3: run a tool-calling loop that mirrors your production workload (search, fetch, summarise, respond, ideally with at least one structured-output step) and count how many of twenty trials complete successfully without manual intervention. Compare to the same loop on your current vendor. If Grok’s success rate is within five percentage points and the cost per loop is lower, the swap is justified for that workload.

Honest caveats

A few things this overview does not pretend to be certain about. The exact rate card on docs.x.ai shifts; the pricing comparisons above are structural (frontier tier vs the now-deprecated cheap-and-fast tier), not numerical, and any specific dollar quote needs a fresh check against the live page. Grok 4.3’s benchmark scores against the latest GPT-5.5 and Claude Sonnet 4.6 checkpoints fluctuate with new releases. xAI’s model-deprecation cadence is also a planning consideration: the May 15 2026 retirement collapsed eight legacy slugs into grok-4.3 in a single window ² , which is a faster deprecation rhythm than OpenAI or Anthropic have run on equivalent tiers, and teams committing engineering time to a specific slug should architect for migration rather than treat the current lineup as fixed.

The X-platform integration assumes you accept the X data-access terms and the privacy implications of your queries hitting xAI infrastructure. For an Indian business subject to DPDP Act 2023 obligations on personal data, double-check that your use case does not put user PII into the Live Search surface in a way that creates a downstream consent problem. Treat any vendor’s latency claim as direction-of-travel rather than gospel; run your own measurements before architecting around them.

The shape of the decision

For most general-purpose developer workloads in 2026, Grok is a strong third or fourth option, not a default. The defaults remain Claude Sonnet 4.6 for code quality, GPT-5.5 for agent reliability, and Gemini 2.5 Pro for long-context cost. After the May 15 2026 retirement, Grok’s wedge is narrower than the pre-retirement framing claimed: X-platform data workflows that no other vendor can serve, very-long-context multi-agent reasoning on Grok 4.20 where the 2-million-token window plus debate runtime genuinely bites, and vendor-diversification strategies where a third frontier in your routing layer reduces single-vendor risk.

That’s a real wedge, and an honest one. Grok in 2026 is a focused tool that wins specific lanes and loses general ones. Pick it for the lanes where it wins, route around it for the rest.